From 243909d41efd266f2c4817bbe968c9ad558df160 Mon Sep 17 00:00:00 2001
From: Robert Bragg <robert@sixbynine.org>
Date: Wed, 25 Nov 2015 16:41:04 +0000
Subject: [PATCH] i965: Add XML OA metric sets for Gen8+

Also updates Makefile.am to generate corresponding normalization code.

Signed-off-by: Robert Bragg <robert@sixbynine.org>
Acked-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
---
 src/mesa/drivers/dri/i965/Makefile.am       |     6 +
 src/mesa/drivers/dri/i965/Makefile.sources  |    14 +-
 src/mesa/drivers/dri/i965/brw_oa_bdw.xml    | 15051 ++++++++++++++++++++++++++
 src/mesa/drivers/dri/i965/brw_oa_bxt.xml    |  9211 ++++++++++++++++
 src/mesa/drivers/dri/i965/brw_oa_chv.xml    |  9569 ++++++++++++++++
 src/mesa/drivers/dri/i965/brw_oa_hsw.xml    |    26 +-
 src/mesa/drivers/dri/i965/brw_oa_sklgt2.xml | 10925 +++++++++++++++++++
 src/mesa/drivers/dri/i965/brw_oa_sklgt3.xml | 10499 ++++++++++++++++++
 src/mesa/drivers/dri/i965/brw_oa_sklgt4.xml | 10522 ++++++++++++++++++
 9 files changed, 65809 insertions(+), 14 deletions(-)
 create mode 100644 src/mesa/drivers/dri/i965/brw_oa_bdw.xml
 create mode 100644 src/mesa/drivers/dri/i965/brw_oa_bxt.xml
 create mode 100644 src/mesa/drivers/dri/i965/brw_oa_chv.xml
 create mode 100644 src/mesa/drivers/dri/i965/brw_oa_sklgt2.xml
 create mode 100644 src/mesa/drivers/dri/i965/brw_oa_sklgt3.xml
 create mode 100644 src/mesa/drivers/dri/i965/brw_oa_sklgt4.xml

diff --git a/src/mesa/drivers/dri/i965/Makefile.am b/src/mesa/drivers/dri/i965/Makefile.am
index fd1695e..03cfe12 100644
--- a/src/mesa/drivers/dri/i965/Makefile.am
+++ b/src/mesa/drivers/dri/i965/Makefile.am
@@ -104,6 +104,12 @@ CLEANFILES = $(BUILT_SOURCES)
 
 EXTRA_DIST = \
 	brw_oa_hsw.xml \
+	brw_oa_bdw.xml \
+	brw_oa_chv.xml \
+	brw_oa_sklgt2.xml \
+	brw_oa_sklgt3.xml \
+	brw_oa_sklgt4.xml \
+	brw_oa_bxt.xml \
 	brw_oa.py
 
 # Note: we avoid using a multi target rule here and outputting both the
diff --git a/src/mesa/drivers/dri/i965/Makefile.sources b/src/mesa/drivers/dri/i965/Makefile.sources
index 405d67f..2905a39 100644
--- a/src/mesa/drivers/dri/i965/Makefile.sources
+++ b/src/mesa/drivers/dri/i965/Makefile.sources
@@ -160,4 +160,16 @@ i965_gen10_FILES = \
 
 i965_oa_GENERATED_FILES = \
 	brw_oa_hsw.h \
-	brw_oa_hsw.c
+	brw_oa_hsw.c \
+	brw_oa_bdw.h \
+	brw_oa_bdw.c \
+	brw_oa_chv.h \
+	brw_oa_chv.c \
+	brw_oa_sklgt2.h \
+	brw_oa_sklgt2.c \
+	brw_oa_sklgt3.h \
+	brw_oa_sklgt3.c \
+	brw_oa_sklgt4.h \
+	brw_oa_sklgt4.c \
+	brw_oa_bxt.h \
+	brw_oa_bxt.c
diff --git a/src/mesa/drivers/dri/i965/brw_oa_bdw.xml b/src/mesa/drivers/dri/i965/brw_oa_bdw.xml
new file mode 100644
index 0000000..ee3c359
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_bdw.xml
@@ -0,0 +1,15051 @@
+<?xml version="1.0"?>
+<metrics version="1491574821" merge_md5="">
+  <set name="Render Metrics Basic Gen8"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="b541bd57-0e0f-4154-b4c0-5858010a2bf7"
+       chipset="BDW"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Cache Misses"
+             description="The total number of sampler cache misses in all LODs in all sampler units."
+             data_type="uint64"
+             equation="B 4 READ B 5 READ UADD 8 UMUL"
+             underscore_name="sampler_l1_misses"
+             units="messages"
+             symbol_name="SamplerL1Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 1 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="$SamplerL1Misses 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="$Sampler0Busy $Sampler1Busy FMAX"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="L3 Lookup Accesses w/o IC"
+             description="The total number of L3 cache lookup accesses w/o IC."
+             data_type="uint64"
+             equation="$SamplerL1Misses $ShaderMemoryAccesses UADD"
+             underscore_name="l3_lookups"
+             units="messages"
+             symbol_name="L3Lookups"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA"
+                     availability="$SliceMask 0x01 AND"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x143F000F" />
+        <register type="NOA" address="0x00009888" value="0x14110014" />
+        <register type="NOA" address="0x00009888" value="0x14310014" />
+        <register type="NOA" address="0x00009888" value="0x14BF000F" />
+        <register type="NOA" address="0x00009888" value="0x118A0317" />
+        <register type="NOA" address="0x00009888" value="0x13837BE0" />
+        <register type="NOA" address="0x00009888" value="0x3B800060" />
+        <register type="NOA" address="0x00009888" value="0x3D800005" />
+        <register type="NOA" address="0x00009888" value="0x005C4000" />
+        <register type="NOA" address="0x00009888" value="0x065C8000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x003D8000" />
+        <register type="NOA" address="0x00009888" value="0x183D0800" />
+        <register type="NOA" address="0x00009888" value="0x0A3F0023" />
+        <register type="NOA" address="0x00009888" value="0x103F0000" />
+        <register type="NOA" address="0x00009888" value="0x00584000" />
+        <register type="NOA" address="0x00009888" value="0x08584000" />
+        <register type="NOA" address="0x00009888" value="0x0A5A4000" />
+        <register type="NOA" address="0x00009888" value="0x005B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B8000" />
+        <register type="NOA" address="0x00009888" value="0x185B2400" />
+        <register type="NOA" address="0x00009888" value="0x0A1D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1F0800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA00" />
+        <register type="NOA" address="0x00009888" value="0x00384000" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18380001" />
+        <register type="NOA" address="0x00009888" value="0x00392000" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A391000" />
+        <register type="NOA" address="0x00009888" value="0x00104000" />
+        <register type="NOA" address="0x00009888" value="0x08104000" />
+        <register type="NOA" address="0x00009888" value="0x00110030" />
+        <register type="NOA" address="0x00009888" value="0x08110031" />
+        <register type="NOA" address="0x00009888" value="0x10110000" />
+        <register type="NOA" address="0x00009888" value="0x00134000" />
+        <register type="NOA" address="0x00009888" value="0x16130020" />
+        <register type="NOA" address="0x00009888" value="0x06308000" />
+        <register type="NOA" address="0x00009888" value="0x08308000" />
+        <register type="NOA" address="0x00009888" value="0x06311800" />
+        <register type="NOA" address="0x00009888" value="0x08311880" />
+        <register type="NOA" address="0x00009888" value="0x10310000" />
+        <register type="NOA" address="0x00009888" value="0x0E334000" />
+        <register type="NOA" address="0x00009888" value="0x16330080" />
+        <register type="NOA" address="0x00009888" value="0x0ABF1180" />
+        <register type="NOA" address="0x00009888" value="0x10BF0000" />
+        <register type="NOA" address="0x00009888" value="0x0ADA8000" />
+        <register type="NOA" address="0x00009888" value="0x0A9D8000" />
+        <register type="NOA" address="0x00009888" value="0x109F0002" />
+        <register type="NOA" address="0x00009888" value="0x0AB94000" />
+        <register type="NOA" address="0x00009888" value="0x0D888000" />
+        <register type="NOA" address="0x00009888" value="0x038A0380" />
+        <register type="NOA" address="0x00009888" value="0x058A000E" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A00A0" />
+        <register type="NOA" address="0x00009888" value="0x078A0000" />
+        <register type="NOA" address="0x00009888" value="0x098A0000" />
+        <register type="NOA" address="0x00009888" value="0x238B2820" />
+        <register type="NOA" address="0x00009888" value="0x258B2550" />
+        <register type="NOA" address="0x00009888" value="0x198C1000" />
+        <register type="NOA" address="0x00009888" value="0x0B8D8000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA0" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x0D831021" />
+        <register type="NOA" address="0x00009888" value="0x0F83572F" />
+        <register type="NOA" address="0x00009888" value="0x01835680" />
+        <register type="NOA" address="0x00009888" value="0x0383002C" />
+        <register type="NOA" address="0x00009888" value="0x11830000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830001" />
+        <register type="NOA" address="0x00009888" value="0x05830000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x05844000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C137" />
+        <register type="NOA" address="0x00009888" value="0x1D80C147" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x15804000" />
+        <register type="NOA" address="0x00009888" value="0x4D801110" />
+        <register type="NOA" address="0x00009888" value="0x4F800331" />
+        <register type="NOA" address="0x00009888" value="0x43800802" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45801465" />
+        <register type="NOA" address="0x00009888" value="0x53801111" />
+        <register type="NOA" address="0x00009888" value="0x478014A5" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800CA5" />
+        <register type="NOA" address="0x00009888" value="0x41800003" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SliceMask 0x02 AND"
+                     priority="1"
+                     >
+        <register type="NOA" address="0x00009888" value="0x143F000F" />
+        <register type="NOA" address="0x00009888" value="0x14BF000F" />
+        <register type="NOA" address="0x00009888" value="0x14910014" />
+        <register type="NOA" address="0x00009888" value="0x14B10014" />
+        <register type="NOA" address="0x00009888" value="0x118A0317" />
+        <register type="NOA" address="0x00009888" value="0x13837BE0" />
+        <register type="NOA" address="0x00009888" value="0x3B800060" />
+        <register type="NOA" address="0x00009888" value="0x3D800005" />
+        <register type="NOA" address="0x00009888" value="0x0A3F0023" />
+        <register type="NOA" address="0x00009888" value="0x103F0000" />
+        <register type="NOA" address="0x00009888" value="0x0A5A4000" />
+        <register type="NOA" address="0x00009888" value="0x0A1D4000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F8000" />
+        <register type="NOA" address="0x00009888" value="0x0A391000" />
+        <register type="NOA" address="0x00009888" value="0x00DC4000" />
+        <register type="NOA" address="0x00009888" value="0x06DC8000" />
+        <register type="NOA" address="0x00009888" value="0x08DCC000" />
+        <register type="NOA" address="0x00009888" value="0x00BD8000" />
+        <register type="NOA" address="0x00009888" value="0x18BD0800" />
+        <register type="NOA" address="0x00009888" value="0x0ABF1180" />
+        <register type="NOA" address="0x00009888" value="0x10BF0000" />
+        <register type="NOA" address="0x00009888" value="0x00D84000" />
+        <register type="NOA" address="0x00009888" value="0x08D84000" />
+        <register type="NOA" address="0x00009888" value="0x0ADA8000" />
+        <register type="NOA" address="0x00009888" value="0x00DB4000" />
+        <register type="NOA" address="0x00009888" value="0x0EDB8000" />
+        <register type="NOA" address="0x00009888" value="0x18DB2400" />
+        <register type="NOA" address="0x00009888" value="0x0A9D8000" />
+        <register type="NOA" address="0x00009888" value="0x0C9F0800" />
+        <register type="NOA" address="0x00009888" value="0x0E9F2A00" />
+        <register type="NOA" address="0x00009888" value="0x109F0002" />
+        <register type="NOA" address="0x00009888" value="0x00B84000" />
+        <register type="NOA" address="0x00009888" value="0x0EB84000" />
+        <register type="NOA" address="0x00009888" value="0x16B84000" />
+        <register type="NOA" address="0x00009888" value="0x18B80001" />
+        <register type="NOA" address="0x00009888" value="0x00B92000" />
+        <register type="NOA" address="0x00009888" value="0x06B98000" />
+        <register type="NOA" address="0x00009888" value="0x08B9A000" />
+        <register type="NOA" address="0x00009888" value="0x0AB94000" />
+        <register type="NOA" address="0x00009888" value="0x00904000" />
+        <register type="NOA" address="0x00009888" value="0x08904000" />
+        <register type="NOA" address="0x00009888" value="0x00910030" />
+        <register type="NOA" address="0x00009888" value="0x08910031" />
+        <register type="NOA" address="0x00009888" value="0x10910000" />
+        <register type="NOA" address="0x00009888" value="0x00934000" />
+        <register type="NOA" address="0x00009888" value="0x16930020" />
+        <register type="NOA" address="0x00009888" value="0x06B08000" />
+        <register type="NOA" address="0x00009888" value="0x08B08000" />
+        <register type="NOA" address="0x00009888" value="0x06B11800" />
+        <register type="NOA" address="0x00009888" value="0x08B11880" />
+        <register type="NOA" address="0x00009888" value="0x10B10000" />
+        <register type="NOA" address="0x00009888" value="0x0EB34000" />
+        <register type="NOA" address="0x00009888" value="0x16B30080" />
+        <register type="NOA" address="0x00009888" value="0x01888000" />
+        <register type="NOA" address="0x00009888" value="0x0D88B800" />
+        <register type="NOA" address="0x00009888" value="0x038A0380" />
+        <register type="NOA" address="0x00009888" value="0x058A000E" />
+        <register type="NOA" address="0x00009888" value="0x1B8A0080" />
+        <register type="NOA" address="0x00009888" value="0x078A0000" />
+        <register type="NOA" address="0x00009888" value="0x098A0000" />
+        <register type="NOA" address="0x00009888" value="0x238B2840" />
+        <register type="NOA" address="0x00009888" value="0x258B26A0" />
+        <register type="NOA" address="0x00009888" value="0x018C4000" />
+        <register type="NOA" address="0x00009888" value="0x0F8C4000" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C1100" />
+        <register type="NOA" address="0x00009888" value="0x018D2000" />
+        <register type="NOA" address="0x00009888" value="0x078D8000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8D8000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA0" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x0D831021" />
+        <register type="NOA" address="0x00009888" value="0x0F83572F" />
+        <register type="NOA" address="0x00009888" value="0x01835680" />
+        <register type="NOA" address="0x00009888" value="0x0383002C" />
+        <register type="NOA" address="0x00009888" value="0x11830000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830001" />
+        <register type="NOA" address="0x00009888" value="0x05830000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x05844000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C137" />
+        <register type="NOA" address="0x00009888" value="0x1D80C147" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x15804000" />
+        <register type="NOA" address="0x00009888" value="0x4D801550" />
+        <register type="NOA" address="0x00009888" value="0x4F800331" />
+        <register type="NOA" address="0x00009888" value="0x43800802" />
+        <register type="NOA" address="0x00009888" value="0x51800400" />
+        <register type="NOA" address="0x00009888" value="0x458004A1" />
+        <register type="NOA" address="0x00009888" value="0x53805555" />
+        <register type="NOA" address="0x00009888" value="0x47800421" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F801421" />
+        <register type="NOA" address="0x00009888" value="0x41800845" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen8"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="35fbc9b2-a891-40a6-a38d-022bb7057552"
+       chipset="BDW"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ C 0 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of typed memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 3 READ B 4 READ B 5 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 1 READ C 2 READ C 3 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ B 2 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL  $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA"
+                     availability="$SliceMask 0x01 AND"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x105C00E0" />
+        <register type="NOA" address="0x00009888" value="0x105800E0" />
+        <register type="NOA" address="0x00009888" value="0x103800E0" />
+        <register type="NOA" address="0x00009888" value="0x3580001A" />
+        <register type="NOA" address="0x00009888" value="0x3B800060" />
+        <register type="NOA" address="0x00009888" value="0x3D800005" />
+        <register type="NOA" address="0x00009888" value="0x065C2100" />
+        <register type="NOA" address="0x00009888" value="0x0A5C0041" />
+        <register type="NOA" address="0x00009888" value="0x0C5C6600" />
+        <register type="NOA" address="0x00009888" value="0x005C6580" />
+        <register type="NOA" address="0x00009888" value="0x085C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5C8000" />
+        <register type="NOA" address="0x00009888" value="0x00580042" />
+        <register type="NOA" address="0x00009888" value="0x08582080" />
+        <register type="NOA" address="0x00009888" value="0x0C58004C" />
+        <register type="NOA" address="0x00009888" value="0x0E582580" />
+        <register type="NOA" address="0x00009888" value="0x005B4000" />
+        <register type="NOA" address="0x00009888" value="0x185B1000" />
+        <register type="NOA" address="0x00009888" value="0x1A5B0104" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA00" />
+        <register type="NOA" address="0x00009888" value="0x101F02AA" />
+        <register type="NOA" address="0x00009888" value="0x08380042" />
+        <register type="NOA" address="0x00009888" value="0x0A382080" />
+        <register type="NOA" address="0x00009888" value="0x0E38404C" />
+        <register type="NOA" address="0x00009888" value="0x0238404B" />
+        <register type="NOA" address="0x00009888" value="0x00384000" />
+        <register type="NOA" address="0x00009888" value="0x16380000" />
+        <register type="NOA" address="0x00009888" value="0x18381145" />
+        <register type="NOA" address="0x00009888" value="0x04380000" />
+        <register type="NOA" address="0x00009888" value="0x0039A000" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x02392000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B02A0" />
+        <register type="NOA" address="0x00009888" value="0x258B5550" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x1F850A80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA0" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x03844000" />
+        <register type="NOA" address="0x00009888" value="0x17808137" />
+        <register type="NOA" address="0x00009888" value="0x1980C147" />
+        <register type="NOA" address="0x00009888" value="0x1B80C0E5" />
+        <register type="NOA" address="0x00009888" value="0x1D80C0E3" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x13804000" />
+        <register type="NOA" address="0x00009888" value="0x15800000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D801000" />
+        <register type="NOA" address="0x00009888" value="0x4F800111" />
+        <register type="NOA" address="0x00009888" value="0x43800062" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800062" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800062" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F801062" />
+        <register type="NOA" address="0x00009888" value="0x41801084" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SliceMask 0x02 AND"
+                     priority="2"
+                     >
+        <register type="NOA" address="0x00009888" value="0x10DC00E0" />
+        <register type="NOA" address="0x00009888" value="0x10D800E0" />
+        <register type="NOA" address="0x00009888" value="0x10B800E0" />
+        <register type="NOA" address="0x00009888" value="0x3580001A" />
+        <register type="NOA" address="0x00009888" value="0x3B800060" />
+        <register type="NOA" address="0x00009888" value="0x3D800005" />
+        <register type="NOA" address="0x00009888" value="0x06DC2100" />
+        <register type="NOA" address="0x00009888" value="0x0ADC0041" />
+        <register type="NOA" address="0x00009888" value="0x0CDC6600" />
+        <register type="NOA" address="0x00009888" value="0x00DC6580" />
+        <register type="NOA" address="0x00009888" value="0x08DC8000" />
+        <register type="NOA" address="0x00009888" value="0x0EDC8000" />
+        <register type="NOA" address="0x00009888" value="0x00D80042" />
+        <register type="NOA" address="0x00009888" value="0x08D82080" />
+        <register type="NOA" address="0x00009888" value="0x0CD8004C" />
+        <register type="NOA" address="0x00009888" value="0x0ED82580" />
+        <register type="NOA" address="0x00009888" value="0x00DB4000" />
+        <register type="NOA" address="0x00009888" value="0x18DB1000" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0104" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E9FAA00" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x08B80042" />
+        <register type="NOA" address="0x00009888" value="0x0AB82080" />
+        <register type="NOA" address="0x00009888" value="0x0EB8404C" />
+        <register type="NOA" address="0x00009888" value="0x02B8404B" />
+        <register type="NOA" address="0x00009888" value="0x00B84000" />
+        <register type="NOA" address="0x00009888" value="0x16B80000" />
+        <register type="NOA" address="0x00009888" value="0x18B81145" />
+        <register type="NOA" address="0x00009888" value="0x04B80000" />
+        <register type="NOA" address="0x00009888" value="0x00B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B98000" />
+        <register type="NOA" address="0x00009888" value="0x08B9A000" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x02B92000" />
+        <register type="NOA" address="0x00009888" value="0x01888000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F800" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x238B0540" />
+        <register type="NOA" address="0x00009888" value="0x258BAAA0" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x018C4000" />
+        <register type="NOA" address="0x00009888" value="0x0F8C4000" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C5500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x018DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D8000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x038D2000" />
+        <register type="NOA" address="0x00009888" value="0x1F850A80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA0" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x03844000" />
+        <register type="NOA" address="0x00009888" value="0x17808137" />
+        <register type="NOA" address="0x00009888" value="0x1980C147" />
+        <register type="NOA" address="0x00009888" value="0x1B80C0E5" />
+        <register type="NOA" address="0x00009888" value="0x1D80C0E3" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x13804000" />
+        <register type="NOA" address="0x00009888" value="0x15800000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D805000" />
+        <register type="NOA" address="0x00009888" value="0x4F800555" />
+        <register type="NOA" address="0x00009888" value="0x43800062" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800062" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800062" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800062" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="233d0544-fff7-4281-8291-e02f222aff72"
+       chipset="BDW"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x0A1E0000" />
+        <register type="NOA" address="0x00009888" value="0x0C1F000F" />
+        <register type="NOA" address="0x00009888" value="0x10176800" />
+        <register type="NOA" address="0x00009888" value="0x1191001F" />
+        <register type="NOA" address="0x00009888" value="0x0B880320" />
+        <register type="NOA" address="0x00009888" value="0x01890C40" />
+        <register type="NOA" address="0x00009888" value="0x118A1C00" />
+        <register type="NOA" address="0x00009888" value="0x118D7C00" />
+        <register type="NOA" address="0x00009888" value="0x118E0020" />
+        <register type="NOA" address="0x00009888" value="0x118F4C00" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x13900001" />
+        <register type="NOA" address="0x00009888" value="0x065C4000" />
+        <register type="NOA" address="0x00009888" value="0x0C3D8000" />
+        <register type="NOA" address="0x00009888" value="0x06584000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B4000" />
+        <register type="NOA" address="0x00009888" value="0x081E0040" />
+        <register type="NOA" address="0x00009888" value="0x0E1E0000" />
+        <register type="NOA" address="0x00009888" value="0x021F5400" />
+        <register type="NOA" address="0x00009888" value="0x001F0000" />
+        <register type="NOA" address="0x00009888" value="0x101F0010" />
+        <register type="NOA" address="0x00009888" value="0x0E1F0080" />
+        <register type="NOA" address="0x00009888" value="0x0C384000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x0C13C000" />
+        <register type="NOA" address="0x00009888" value="0x06164000" />
+        <register type="NOA" address="0x00009888" value="0x06170012" />
+        <register type="NOA" address="0x00009888" value="0x00170000" />
+        <register type="NOA" address="0x00009888" value="0x01910005" />
+        <register type="NOA" address="0x00009888" value="0x07880002" />
+        <register type="NOA" address="0x00009888" value="0x01880C00" />
+        <register type="NOA" address="0x00009888" value="0x0F880000" />
+        <register type="NOA" address="0x00009888" value="0x0D880000" />
+        <register type="NOA" address="0x00009888" value="0x05880000" />
+        <register type="NOA" address="0x00009888" value="0x09890032" />
+        <register type="NOA" address="0x00009888" value="0x078A0800" />
+        <register type="NOA" address="0x00009888" value="0x0F8A0A00" />
+        <register type="NOA" address="0x00009888" value="0x198A4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A2000" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0000" />
+        <register type="NOA" address="0x00009888" value="0x038A4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B54C0" />
+        <register type="NOA" address="0x00009888" value="0x258BAA55" />
+        <register type="NOA" address="0x00009888" value="0x278B0019" />
+        <register type="NOA" address="0x00009888" value="0x198C0100" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x0F8D0015" />
+        <register type="NOA" address="0x00009888" value="0x018D1000" />
+        <register type="NOA" address="0x00009888" value="0x098D8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DF000" />
+        <register type="NOA" address="0x00009888" value="0x0D8D3000" />
+        <register type="NOA" address="0x00009888" value="0x038DE000" />
+        <register type="NOA" address="0x00009888" value="0x058D3000" />
+        <register type="NOA" address="0x00009888" value="0x0D8E0004" />
+        <register type="NOA" address="0x00009888" value="0x058E000C" />
+        <register type="NOA" address="0x00009888" value="0x098E0000" />
+        <register type="NOA" address="0x00009888" value="0x078E0000" />
+        <register type="NOA" address="0x00009888" value="0x038E0000" />
+        <register type="NOA" address="0x00009888" value="0x0B8F0020" />
+        <register type="NOA" address="0x00009888" value="0x198F0C00" />
+        <register type="NOA" address="0x00009888" value="0x078F8000" />
+        <register type="NOA" address="0x00009888" value="0x098F4000" />
+        <register type="NOA" address="0x00009888" value="0x0B900980" />
+        <register type="NOA" address="0x00009888" value="0x03900D80" />
+        <register type="NOA" address="0x00009888" value="0x01900000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D801111" />
+        <register type="NOA" address="0x00009888" value="0x3D800800" />
+        <register type="NOA" address="0x00009888" value="0x4F801011" />
+        <register type="NOA" address="0x00009888" value="0x43800443" />
+        <register type="NOA" address="0x00009888" value="0x51801111" />
+        <register type="NOA" address="0x00009888" value="0x45800422" />
+        <register type="NOA" address="0x00009888" value="0x53801111" />
+        <register type="NOA" address="0x00009888" value="0x47800C60" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800422" />
+        <register type="NOA" address="0x00009888" value="0x41800021" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen8"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="2b255d48-2117-4fef-a8f7-f151e1d25a2c"
+       chipset="BDW"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank0Reads"
+             description="The total number of GTI memory reads from L3 Bank 0 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_reads"
+             units="messages"
+             symbol_name="GtiL3Bank0Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all accesses from GTI to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiL3Bank3Reads"
+             description="The total number of GTI memory reads from L3 Bank 3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_reads"
+             units="messages"
+             symbol_name="GtiL3Bank3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiRsMemoryReads"
+             description="The total number of GTI memory reads from Resource Streamer."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_rs_memory_reads"
+             units="messages"
+             symbol_name="GtiRsMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiHizMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache (Hi-Depth Cache misses)."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_reads"
+             units="messages"
+             symbol_name="GtiHizMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="GtiRccMemoryReads"
+             description="The total number of GTI memory reads from Render Color Cache (Render Color Cache misses)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_reads"
+             units="messages"
+             symbol_name="GtiRccMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank1Reads"
+             description="The total number of GTI memory reads from L3 Bank 1 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_reads"
+             units="messages"
+             symbol_name="GtiL3Bank1Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiCmdStreamerMemoryReads"
+             description="The total number of GTI memory reads from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_reads"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GtiL3Bank2Reads"
+             description="The total number of GTI memory reads from L3 Bank 2 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_reads"
+             units="messages"
+             symbol_name="GtiL3Bank2Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiMemoryReads"
+             description="The total number of GTI memory reads."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache (Render Depth Cache misses)."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="GtiMscMemoryReads"
+             description="The total number of GTI memory reads from Multisampling Color Cache (Multisampling Color Cache misses)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_reads"
+             units="messages"
+             symbol_name="GtiMscMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiVfMemoryReads"
+             description="The total number of GTI memory reads from Vertex Fetch."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="gti_vf_memory_reads"
+             units="messages"
+             symbol_name="GtiVfMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache (Stencil Cache misses)."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Reads $GtiL3Bank1Reads $GtiL3Bank2Reads $GtiL3Bank3Reads UADD UADD UADD"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x198B0343" />
+        <register type="NOA" address="0x00009888" value="0x13845800" />
+        <register type="NOA" address="0x00009888" value="0x15840018" />
+        <register type="NOA" address="0x00009888" value="0x3580001A" />
+        <register type="NOA" address="0x00009888" value="0x038B6300" />
+        <register type="NOA" address="0x00009888" value="0x058B6B62" />
+        <register type="NOA" address="0x00009888" value="0x078B006A" />
+        <register type="NOA" address="0x00009888" value="0x118B0000" />
+        <register type="NOA" address="0x00009888" value="0x238B0000" />
+        <register type="NOA" address="0x00009888" value="0x258B0000" />
+        <register type="NOA" address="0x00009888" value="0x1F85A080" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385000A" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x01840018" />
+        <register type="NOA" address="0x00009888" value="0x07844C80" />
+        <register type="NOA" address="0x00009888" value="0x09840D9A" />
+        <register type="NOA" address="0x00009888" value="0x0B840E9C" />
+        <register type="NOA" address="0x00009888" value="0x0D840F9E" />
+        <register type="NOA" address="0x00009888" value="0x0F840010" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x03848000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x2F8000E5" />
+        <register type="NOA" address="0x00009888" value="0x138080E3" />
+        <register type="NOA" address="0x00009888" value="0x1580C0E1" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x11804000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F804000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800800" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800842" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800842" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47801042" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800084" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F872" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen8"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="f7fd3220-b466-4a4d-9f98-b0caf3f2394c"
+       chipset="BDW"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiMemoryWrites"
+             description="The total number of GTI memory writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all GTI accesses to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiMscMemoryWrites"
+             description="The total number of GTI memory writes from Multisampling Color Cache (Multisampling Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_writes"
+             units="messages"
+             symbol_name="GtiMscMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiCmdStreamerMemoryWrites"
+             description="The total number of GTI memory writes from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_writes"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiL3Bank0Writes"
+             description="The total number of GTI memory writes from L3 Bank 0 (L3 Bank 0 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_writes"
+             units="messages"
+             symbol_name="GtiL3Bank0Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank1Writes"
+             description="The total number of GTI memory writes from L3 Bank 1 (L3 Bank 1 invalidations)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_writes"
+             units="messages"
+             symbol_name="GtiL3Bank1Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank2Writes"
+             description="The total number of GTI memory writes from L3 Bank 2 (L3 Bank 2 invalidations)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_writes"
+             units="messages"
+             symbol_name="GtiL3Bank2Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank3Writes"
+             description="The total number of GTI memory writes from L3 Bank 3 (L3 Bank 3 invalidations)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_writes"
+             units="messages"
+             symbol_name="GtiL3Bank3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Writes $GtiL3Bank1Writes $GtiL3Bank2Writes $GtiL3Bank3Writes UADD UADD UADD"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRccMemoryWrites"
+             description="The total number of GTI memory writes from Render Color Cache (Render Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_writes"
+             units="messages"
+             symbol_name="GtiRccMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiSoMemoryWrites"
+             description="The total number of GTI memory writes from Stream Output."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_so_memory_writes"
+             units="messages"
+             symbol_name="GtiSoMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiStcMemoryWrites"
+             description="The total number of GTI memory writes from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_writes"
+             units="messages"
+             symbol_name="GtiStcMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GtiRczMemoryWrites"
+             description="The total number of GTI memory writes from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_writes"
+             units="messages"
+             symbol_name="GtiRczMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiHizMemoryWrites"
+             description="The total number of GTI memory writes from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_writes"
+             units="messages"
+             symbol_name="GtiHizMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x198B0343" />
+        <register type="NOA" address="0x00009888" value="0x13845400" />
+        <register type="NOA" address="0x00009888" value="0x3580001A" />
+        <register type="NOA" address="0x00009888" value="0x3D800805" />
+        <register type="NOA" address="0x00009888" value="0x038B6300" />
+        <register type="NOA" address="0x00009888" value="0x058B6B62" />
+        <register type="NOA" address="0x00009888" value="0x078B006A" />
+        <register type="NOA" address="0x00009888" value="0x118B0000" />
+        <register type="NOA" address="0x00009888" value="0x238B0000" />
+        <register type="NOA" address="0x00009888" value="0x258B0000" />
+        <register type="NOA" address="0x00009888" value="0x1F85A080" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x23850002" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x01840010" />
+        <register type="NOA" address="0x00009888" value="0x07844880" />
+        <register type="NOA" address="0x00009888" value="0x09840992" />
+        <register type="NOA" address="0x00009888" value="0x0B840A94" />
+        <register type="NOA" address="0x00009888" value="0x0D840B96" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x03848000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x2D800147" />
+        <register type="NOA" address="0x00009888" value="0x2F8000E5" />
+        <register type="NOA" address="0x00009888" value="0x138080E3" />
+        <register type="NOA" address="0x00009888" value="0x1580C0E1" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x11804000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F800000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800842" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800842" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47801082" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800084" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F822" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen8"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="e99ccaca-821c-4df9-97a7-96bdb7204e43"
+       chipset="BDW"
+       symbol_name="ComputeExtended"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Writes 0"
+             description="The subslice 0 typed writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="typed_writes0"
+             units="messages"
+             symbol_name="TypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuTypedAtomics0"
+             description="The subslice 0 EU Typed Atomics subslice 0."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_typed_atomics0"
+             units="messages"
+             symbol_name="EuTypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Atomics 0"
+             description="The subslice 0 typed atomics."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="typed_atomics0"
+             units="messages"
+             symbol_name="TypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedAtomicsPerCacheLine"
+             description="Ratio of EU typed atomics requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedAtomics0 $TypedAtomics0 FDIV"
+             underscore_name="typed_atomics_per_cache_line"
+             units="eu atomic requests to l3 cache lines"
+             symbol_name="TypedAtomicsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedReads0"
+             description="The subslice 0 EU Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_untyped_reads0"
+             units="messages"
+             symbol_name="EuUntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Writes 0"
+             description="The subslice 0 untyped writes (including SLM writes)."
+             data_type="uint64"
+             equation="C 1 READ"
+             underscore_name="untyped_writes0"
+             units="messages"
+             symbol_name="UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedAtomics0"
+             description="The subslice 0 EU Untyped Atomics subslice 0."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_untyped_atomics0"
+             units="messages"
+             symbol_name="EuUntypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedWrites0"
+             description="The subslice 0 EU Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="eu_untyped_writes0"
+             units="messages"
+             symbol_name="EuUntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedWrites0"
+             description="The subslice 0 EU A64 Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_a64_untyped_writes0"
+             units="messages"
+             symbol_name="EuA64UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedWritesPerCacheLine"
+             description="Ratio of EU untyped write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuUntypedWrites0 $EuA64UntypedWrites0 UADD $UntypedWrites0 FDIV"
+             underscore_name="untyped_writes_per_cache_line"
+             units="eu requests to l3 cache lines"
+             symbol_name="UntypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedWrites0"
+             description="The subslice 0 EU Typed Writes subslice 0."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="eu_typed_writes0"
+             units="messages"
+             symbol_name="EuTypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedWritesPerCacheLine"
+             description="Ratio of EU typed write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedWrites0 $TypedWrites0 FDIV"
+             underscore_name="typed_writes_per_cache_line"
+             units="eu requests to l3 cache lines"
+             symbol_name="TypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Reads 0"
+             description="The subslice 0 typed reads."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="typed_reads0"
+             units="messages"
+             symbol_name="TypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Reads 0"
+             description="The subslice 0 untyped reads (including SLM reads)."
+             data_type="uint64"
+             equation="C 3 READ"
+             underscore_name="untyped_reads0"
+             units="messages"
+             symbol_name="UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedReads0"
+             description="The subslice 0 EU A64 Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_a64_untyped_reads0"
+             units="messages"
+             symbol_name="EuA64UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedReads0"
+             description="The subslice 0 EU Typed Reads subslice 0."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="eu_typed_reads0"
+             units="messages"
+             symbol_name="EuTypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedReadsPerCacheLine"
+             description="Ratio of EU untyped read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuUntypedReads0 $EuA64UntypedReads0 UADD $UntypedReads0 FDIV"
+             underscore_name="untyped_reads_per_cache_line"
+             units="eu requests to l3 cache lines"
+             symbol_name="UntypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="Ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuTypedReads0 $TypedReads0 FDIV"
+             underscore_name="typed_reads_per_cache_line"
+             units="eu requests to l3 cache lines"
+             symbol_name="TypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x01 AND"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x143D0160" />
+        <register type="NOA" address="0x00009888" value="0x163D2800" />
+        <register type="NOA" address="0x00009888" value="0x183D0120" />
+        <register type="NOA" address="0x00009888" value="0x105800E0" />
+        <register type="NOA" address="0x00009888" value="0x005CC000" />
+        <register type="NOA" address="0x00009888" value="0x065C8000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0E5CC000" />
+        <register type="NOA" address="0x00009888" value="0x025CC000" />
+        <register type="NOA" address="0x00009888" value="0x045CC000" />
+        <register type="NOA" address="0x00009888" value="0x003D0011" />
+        <register type="NOA" address="0x00009888" value="0x063D0900" />
+        <register type="NOA" address="0x00009888" value="0x083D0A13" />
+        <register type="NOA" address="0x00009888" value="0x0A3D0B15" />
+        <register type="NOA" address="0x00009888" value="0x0C3D2317" />
+        <register type="NOA" address="0x00009888" value="0x043D21B7" />
+        <register type="NOA" address="0x00009888" value="0x103D0000" />
+        <register type="NOA" address="0x00009888" value="0x0E3D0000" />
+        <register type="NOA" address="0x00009888" value="0x1A3D0000" />
+        <register type="NOA" address="0x00009888" value="0x0E5825C1" />
+        <register type="NOA" address="0x00009888" value="0x00586100" />
+        <register type="NOA" address="0x00009888" value="0x0258204C" />
+        <register type="NOA" address="0x00009888" value="0x06588000" />
+        <register type="NOA" address="0x00009888" value="0x0858C000" />
+        <register type="NOA" address="0x00009888" value="0x0A58C000" />
+        <register type="NOA" address="0x00009888" value="0x0C58C000" />
+        <register type="NOA" address="0x00009888" value="0x0458C000" />
+        <register type="NOA" address="0x00009888" value="0x005B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x185B5400" />
+        <register type="NOA" address="0x00009888" value="0x1A5B0155" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA2A" />
+        <register type="NOA" address="0x00009888" value="0x101F02AA" />
+        <register type="NOA" address="0x00009888" value="0x00384000" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18381555" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x04384000" />
+        <register type="NOA" address="0x00009888" value="0x06384000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0039A000" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B2AA0" />
+        <register type="NOA" address="0x00009888" value="0x258B5551" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA2" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x02 AND"
+                     priority="2"
+                     >
+        <register type="NOA" address="0x00009888" value="0x105C00E0" />
+        <register type="NOA" address="0x00009888" value="0x145B0160" />
+        <register type="NOA" address="0x00009888" value="0x165B2800" />
+        <register type="NOA" address="0x00009888" value="0x185B0120" />
+        <register type="NOA" address="0x00009888" value="0x0E5C25C1" />
+        <register type="NOA" address="0x00009888" value="0x005C6100" />
+        <register type="NOA" address="0x00009888" value="0x025C204C" />
+        <register type="NOA" address="0x00009888" value="0x065C8000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5CC000" />
+        <register type="NOA" address="0x00009888" value="0x045CC000" />
+        <register type="NOA" address="0x00009888" value="0x005B0011" />
+        <register type="NOA" address="0x00009888" value="0x065B0900" />
+        <register type="NOA" address="0x00009888" value="0x085B0A13" />
+        <register type="NOA" address="0x00009888" value="0x0A5B0B15" />
+        <register type="NOA" address="0x00009888" value="0x0C5B2317" />
+        <register type="NOA" address="0x00009888" value="0x045B21B7" />
+        <register type="NOA" address="0x00009888" value="0x105B0000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B0000" />
+        <register type="NOA" address="0x00009888" value="0x1A5B0000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA2A" />
+        <register type="NOA" address="0x00009888" value="0x101F02AA" />
+        <register type="NOA" address="0x00009888" value="0x00384000" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18381555" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x04384000" />
+        <register type="NOA" address="0x00009888" value="0x06384000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0039A000" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B2AA0" />
+        <register type="NOA" address="0x00009888" value="0x258B5551" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA2" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x04 AND"
+                     priority="4"
+                     >
+        <register type="NOA" address="0x00009888" value="0x103800E0" />
+        <register type="NOA" address="0x00009888" value="0x143A0160" />
+        <register type="NOA" address="0x00009888" value="0x163A2800" />
+        <register type="NOA" address="0x00009888" value="0x183A0120" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA2A" />
+        <register type="NOA" address="0x00009888" value="0x101F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0E38A5C1" />
+        <register type="NOA" address="0x00009888" value="0x0038A100" />
+        <register type="NOA" address="0x00009888" value="0x0238204C" />
+        <register type="NOA" address="0x00009888" value="0x16388000" />
+        <register type="NOA" address="0x00009888" value="0x183802AA" />
+        <register type="NOA" address="0x00009888" value="0x04380000" />
+        <register type="NOA" address="0x00009888" value="0x06380000" />
+        <register type="NOA" address="0x00009888" value="0x08388000" />
+        <register type="NOA" address="0x00009888" value="0x0A388000" />
+        <register type="NOA" address="0x00009888" value="0x0039A000" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x003A0011" />
+        <register type="NOA" address="0x00009888" value="0x063A0900" />
+        <register type="NOA" address="0x00009888" value="0x083A0A13" />
+        <register type="NOA" address="0x00009888" value="0x0A3A0B15" />
+        <register type="NOA" address="0x00009888" value="0x0C3A2317" />
+        <register type="NOA" address="0x00009888" value="0x043A21B7" />
+        <register type="NOA" address="0x00009888" value="0x103A0000" />
+        <register type="NOA" address="0x00009888" value="0x0E3A0000" />
+        <register type="NOA" address="0x00009888" value="0x1A3A0000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B2AA0" />
+        <register type="NOA" address="0x00009888" value="0x258B5551" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA2" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x08 AND"
+                     priority="1"
+                     >
+        <register type="NOA" address="0x00009888" value="0x14BD0160" />
+        <register type="NOA" address="0x00009888" value="0x16BD2800" />
+        <register type="NOA" address="0x00009888" value="0x18BD0120" />
+        <register type="NOA" address="0x00009888" value="0x10D800E0" />
+        <register type="NOA" address="0x00009888" value="0x00DCC000" />
+        <register type="NOA" address="0x00009888" value="0x06DC8000" />
+        <register type="NOA" address="0x00009888" value="0x08DCC000" />
+        <register type="NOA" address="0x00009888" value="0x0ADCC000" />
+        <register type="NOA" address="0x00009888" value="0x0CDCC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDCC000" />
+        <register type="NOA" address="0x00009888" value="0x02DCC000" />
+        <register type="NOA" address="0x00009888" value="0x04DCC000" />
+        <register type="NOA" address="0x00009888" value="0x00BD0011" />
+        <register type="NOA" address="0x00009888" value="0x06BD0900" />
+        <register type="NOA" address="0x00009888" value="0x08BD0A13" />
+        <register type="NOA" address="0x00009888" value="0x0ABD0B15" />
+        <register type="NOA" address="0x00009888" value="0x0CBD2317" />
+        <register type="NOA" address="0x00009888" value="0x04BD21B7" />
+        <register type="NOA" address="0x00009888" value="0x10BD0000" />
+        <register type="NOA" address="0x00009888" value="0x0EBD0000" />
+        <register type="NOA" address="0x00009888" value="0x1ABD0000" />
+        <register type="NOA" address="0x00009888" value="0x0ED825C1" />
+        <register type="NOA" address="0x00009888" value="0x00D86100" />
+        <register type="NOA" address="0x00009888" value="0x02D8204C" />
+        <register type="NOA" address="0x00009888" value="0x06D88000" />
+        <register type="NOA" address="0x00009888" value="0x08D8C000" />
+        <register type="NOA" address="0x00009888" value="0x0AD8C000" />
+        <register type="NOA" address="0x00009888" value="0x0CD8C000" />
+        <register type="NOA" address="0x00009888" value="0x04D8C000" />
+        <register type="NOA" address="0x00009888" value="0x00DB4000" />
+        <register type="NOA" address="0x00009888" value="0x0EDB4000" />
+        <register type="NOA" address="0x00009888" value="0x18DB5400" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0155" />
+        <register type="NOA" address="0x00009888" value="0x02DB4000" />
+        <register type="NOA" address="0x00009888" value="0x04DB4000" />
+        <register type="NOA" address="0x00009888" value="0x06DB4000" />
+        <register type="NOA" address="0x00009888" value="0x08DB4000" />
+        <register type="NOA" address="0x00009888" value="0x0ADB4000" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E9FAA2A" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x00B84000" />
+        <register type="NOA" address="0x00009888" value="0x0EB84000" />
+        <register type="NOA" address="0x00009888" value="0x16B84000" />
+        <register type="NOA" address="0x00009888" value="0x18B81555" />
+        <register type="NOA" address="0x00009888" value="0x02B84000" />
+        <register type="NOA" address="0x00009888" value="0x04B84000" />
+        <register type="NOA" address="0x00009888" value="0x06B84000" />
+        <register type="NOA" address="0x00009888" value="0x08B84000" />
+        <register type="NOA" address="0x00009888" value="0x0AB84000" />
+        <register type="NOA" address="0x00009888" value="0x00B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B98000" />
+        <register type="NOA" address="0x00009888" value="0x08B9A000" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x04B9A000" />
+        <register type="NOA" address="0x00009888" value="0x01888000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F800" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x238B5540" />
+        <register type="NOA" address="0x00009888" value="0x258BAAA2" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x018C4000" />
+        <register type="NOA" address="0x00009888" value="0x0F8C4000" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C5500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x018DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D8000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA2" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x10 AND"
+                     priority="3"
+                     >
+        <register type="NOA" address="0x00009888" value="0x10DC00E0" />
+        <register type="NOA" address="0x00009888" value="0x14DB0160" />
+        <register type="NOA" address="0x00009888" value="0x16DB2800" />
+        <register type="NOA" address="0x00009888" value="0x18DB0120" />
+        <register type="NOA" address="0x00009888" value="0x0EDC25C1" />
+        <register type="NOA" address="0x00009888" value="0x00DC6100" />
+        <register type="NOA" address="0x00009888" value="0x02DC204C" />
+        <register type="NOA" address="0x00009888" value="0x06DC8000" />
+        <register type="NOA" address="0x00009888" value="0x08DCC000" />
+        <register type="NOA" address="0x00009888" value="0x0ADCC000" />
+        <register type="NOA" address="0x00009888" value="0x0CDCC000" />
+        <register type="NOA" address="0x00009888" value="0x04DCC000" />
+        <register type="NOA" address="0x00009888" value="0x00DB0011" />
+        <register type="NOA" address="0x00009888" value="0x06DB0900" />
+        <register type="NOA" address="0x00009888" value="0x08DB0A13" />
+        <register type="NOA" address="0x00009888" value="0x0ADB0B15" />
+        <register type="NOA" address="0x00009888" value="0x0CDB2317" />
+        <register type="NOA" address="0x00009888" value="0x04DB21B7" />
+        <register type="NOA" address="0x00009888" value="0x10DB0000" />
+        <register type="NOA" address="0x00009888" value="0x0EDB0000" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0000" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E9FAA2A" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x00B84000" />
+        <register type="NOA" address="0x00009888" value="0x0EB84000" />
+        <register type="NOA" address="0x00009888" value="0x16B84000" />
+        <register type="NOA" address="0x00009888" value="0x18B81555" />
+        <register type="NOA" address="0x00009888" value="0x02B84000" />
+        <register type="NOA" address="0x00009888" value="0x04B84000" />
+        <register type="NOA" address="0x00009888" value="0x06B84000" />
+        <register type="NOA" address="0x00009888" value="0x08B84000" />
+        <register type="NOA" address="0x00009888" value="0x0AB84000" />
+        <register type="NOA" address="0x00009888" value="0x00B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B98000" />
+        <register type="NOA" address="0x00009888" value="0x08B9A000" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x04B9A000" />
+        <register type="NOA" address="0x00009888" value="0x01888000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F800" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x238B5540" />
+        <register type="NOA" address="0x00009888" value="0x258BAAA2" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x018C4000" />
+        <register type="NOA" address="0x00009888" value="0x0F8C4000" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C5500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x018DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D8000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA2" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x20 AND"
+                     priority="5"
+                     >
+        <register type="NOA" address="0x00009888" value="0x10B800E0" />
+        <register type="NOA" address="0x00009888" value="0x14BA0160" />
+        <register type="NOA" address="0x00009888" value="0x16BA2800" />
+        <register type="NOA" address="0x00009888" value="0x18BA0120" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E9FAA2A" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0EB8A5C1" />
+        <register type="NOA" address="0x00009888" value="0x00B8A100" />
+        <register type="NOA" address="0x00009888" value="0x02B8204C" />
+        <register type="NOA" address="0x00009888" value="0x16B88000" />
+        <register type="NOA" address="0x00009888" value="0x18B802AA" />
+        <register type="NOA" address="0x00009888" value="0x04B80000" />
+        <register type="NOA" address="0x00009888" value="0x06B80000" />
+        <register type="NOA" address="0x00009888" value="0x08B88000" />
+        <register type="NOA" address="0x00009888" value="0x0AB88000" />
+        <register type="NOA" address="0x00009888" value="0x00B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B98000" />
+        <register type="NOA" address="0x00009888" value="0x08B9A000" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x04B9A000" />
+        <register type="NOA" address="0x00009888" value="0x00BA0011" />
+        <register type="NOA" address="0x00009888" value="0x06BA0900" />
+        <register type="NOA" address="0x00009888" value="0x08BA0A13" />
+        <register type="NOA" address="0x00009888" value="0x0ABA0B15" />
+        <register type="NOA" address="0x00009888" value="0x0CBA2317" />
+        <register type="NOA" address="0x00009888" value="0x04BA21B7" />
+        <register type="NOA" address="0x00009888" value="0x10BA0000" />
+        <register type="NOA" address="0x00009888" value="0x0EBA0000" />
+        <register type="NOA" address="0x00009888" value="0x1ABA0000" />
+        <register type="NOA" address="0x00009888" value="0x01888000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F800" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x238B5540" />
+        <register type="NOA" address="0x00009888" value="0x258BAAA2" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x018C4000" />
+        <register type="NOA" address="0x00009888" value="0x0F8C4000" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C5500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x018DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D8000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA2" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FC2A" />
+        <register type="OA" address="0x00002774" value="0x0000BF00" />
+        <register type="OA" address="0x00002778" value="0x0007FC6A" />
+        <register type="OA" address="0x0000277C" value="0x0000BF00" />
+        <register type="OA" address="0x00002780" value="0x0007FC92" />
+        <register type="OA" address="0x00002784" value="0x0000BF00" />
+        <register type="OA" address="0x00002788" value="0x0007FCA2" />
+        <register type="OA" address="0x0000278C" value="0x0000BF00" />
+        <register type="OA" address="0x00002790" value="0x0007FC32" />
+        <register type="OA" address="0x00002794" value="0x0000BF00" />
+        <register type="OA" address="0x00002798" value="0x0007FC9A" />
+        <register type="OA" address="0x0000279C" value="0x0000BF00" />
+        <register type="OA" address="0x000027A0" value="0x0007FE6A" />
+        <register type="OA" address="0x000027A4" value="0x0000BF00" />
+        <register type="OA" address="0x000027A8" value="0x0007FE7A" />
+        <register type="OA" address="0x000027AC" value="0x0000BF00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics L3 Cache Gen8"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_l3_cache"
+       hw_config_guid="27a364dc-8225-4ecb-b607-d6f1925598d9"
+       chipset="BDW"
+       symbol_name="ComputeL3Cache"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 03 Accesses"
+             description="The total number of accesses to L3 Bank 03."
+             data_type="uint64"
+             equation="B 3 READ 2 UMUL"
+             underscore_name="l3_bank03_accesses"
+             units="messages"
+             symbol_name="L3Bank03Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Accesses"
+             description="The total number of L3 accesses from all entities."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ B 2 READ B 3 READ C 2 READ C 3 READ B 6 READ B 7 READ UADD UADD UADD UADD UADD UADD UADD 2 UMUL"
+             underscore_name="l3_accesses"
+             units="messages"
+             symbol_name="L3Accesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Total Throughput"
+             description="The total number of GPU memory bytes transferred via L3."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Accesses 64 UMUL"
+             underscore_name="l3_total_throughput"
+             units="bytes"
+             symbol_name="L3TotalThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OCL OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OCL OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="EU FPU0 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu0_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ C 5 READ UADD"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="L3 Bank 00 Accesses"
+             description="The total number of accesses to L3 Bank 00."
+             data_type="uint64"
+             equation="C 0 READ 2 UMUL"
+             underscore_name="l3_bank00_accesses"
+             units="messages"
+             symbol_name="L3Bank00Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU0 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 19 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu0_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="L3 Bank 10 Accesses"
+             description="The total number of accesses to L3 Bank 10."
+             data_type="uint64"
+             equation="C 2 READ 2 UMUL"
+             underscore_name="l3_bank10_accesses"
+             units="messages"
+             symbol_name="L3Bank10Accesses"
+             availability="$SliceMask 0x02 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU1 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 14 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu1_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 10 IC Accesses"
+             description="The total number of accesses to L3 Bank 10 from IC cache."
+             data_type="uint64"
+             equation="B 4 READ B 5 READ UADD 2 UMUL $L3Bank10Accesses UMIN"
+             underscore_name="l3_bank10_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank10IcAccesses"
+             availability="$SliceMask 0x02 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 10 IC Hits"
+             description="The total number of hits in L3 Bank 10 from IC cache."
+             data_type="uint64"
+             equation="B 5 READ 2 UMUL $L3Bank10IcAccesses UMIN"
+             underscore_name="l3_bank10_ic_hits"
+             units="messages"
+             symbol_name="L3Bank10IcHits"
+             availability="$SliceMask 0x02 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 00 IC Accesses"
+             description="The total number of accesses to L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 0 READ B 1 READ UADD 2 UMUL $L3Bank00Accesses UMIN"
+             underscore_name="l3_bank00_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank00IcAccesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 00 IC Hits"
+             description="The total number of hits in L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 1 READ 2 UMUL $L3Bank00IcAccesses UMIN"
+             underscore_name="l3_bank00_ic_hits"
+             units="messages"
+             symbol_name="L3Bank00IcHits"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 01 Accesses"
+             description="The total number of accesses to L3 Bank 01."
+             data_type="uint64"
+             equation="C 1 READ 2 UMUL"
+             underscore_name="l3_bank01_accesses"
+             units="messages"
+             symbol_name="L3Bank01Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 20 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu1_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="L3 Bank 11 Accesses"
+             description="The total number of accesses to L3 Bank 11."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="l3_bank11_accesses"
+             units="messages"
+             symbol_name="L3Bank11Accesses"
+             availability="$SliceMask 0x02 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Bank 02 Accesses"
+             description="The total number of accesses to L3 Bank 02."
+             data_type="uint64"
+             equation="B 2 READ 2 UMUL"
+             underscore_name="l3_bank02_accesses"
+             units="messages"
+             symbol_name="L3Bank02Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Bank 13 Accesses"
+             description="The total number of accesses to L3 Bank 13."
+             data_type="uint64"
+             equation="B 7 READ 2 UMUL"
+             underscore_name="l3_bank13_accesses"
+             units="messages"
+             symbol_name="L3Bank13Accesses"
+             availability="$SliceMask 0x02 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="L3 Bank 12 Accesses"
+             description="The total number of accesses to L3 Bank 12."
+             data_type="uint64"
+             equation="B 6 READ 2 UMUL"
+             underscore_name="l3_bank12_accesses"
+             units="messages"
+             symbol_name="L3Bank12Accesses"
+             availability="$SliceMask 0x02 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU1 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x143F00B3" />
+        <register type="NOA" address="0x00009888" value="0x14BF00B3" />
+        <register type="NOA" address="0x00009888" value="0x138303C0" />
+        <register type="NOA" address="0x00009888" value="0x3B800060" />
+        <register type="NOA" address="0x00009888" value="0x3D800805" />
+        <register type="NOA" address="0x00009888" value="0x003F0029" />
+        <register type="NOA" address="0x00009888" value="0x063F1400" />
+        <register type="NOA" address="0x00009888" value="0x083F1225" />
+        <register type="NOA" address="0x00009888" value="0x0E3F1327" />
+        <register type="NOA" address="0x00009888" value="0x103F0000" />
+        <register type="NOA" address="0x00009888" value="0x005A4000" />
+        <register type="NOA" address="0x00009888" value="0x065A8000" />
+        <register type="NOA" address="0x00009888" value="0x085AC000" />
+        <register type="NOA" address="0x00009888" value="0x0E5AC000" />
+        <register type="NOA" address="0x00009888" value="0x001D4000" />
+        <register type="NOA" address="0x00009888" value="0x061D8000" />
+        <register type="NOA" address="0x00009888" value="0x081DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1DC000" />
+        <register type="NOA" address="0x00009888" value="0x0C1F0800" />
+        <register type="NOA" address="0x00009888" value="0x0E1F2A00" />
+        <register type="NOA" address="0x00009888" value="0x101F0280" />
+        <register type="NOA" address="0x00009888" value="0x00391000" />
+        <register type="NOA" address="0x00009888" value="0x06394000" />
+        <register type="NOA" address="0x00009888" value="0x08395000" />
+        <register type="NOA" address="0x00009888" value="0x0E395000" />
+        <register type="NOA" address="0x00009888" value="0x0ABF1429" />
+        <register type="NOA" address="0x00009888" value="0x0CBF1225" />
+        <register type="NOA" address="0x00009888" value="0x00BF1380" />
+        <register type="NOA" address="0x00009888" value="0x02BF0026" />
+        <register type="NOA" address="0x00009888" value="0x10BF0000" />
+        <register type="NOA" address="0x00009888" value="0x0ADAC000" />
+        <register type="NOA" address="0x00009888" value="0x0CDAC000" />
+        <register type="NOA" address="0x00009888" value="0x00DA8000" />
+        <register type="NOA" address="0x00009888" value="0x02DA4000" />
+        <register type="NOA" address="0x00009888" value="0x0A9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0C9DC000" />
+        <register type="NOA" address="0x00009888" value="0x009D8000" />
+        <register type="NOA" address="0x00009888" value="0x029D4000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F8000" />
+        <register type="NOA" address="0x00009888" value="0x109F002A" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0AB95000" />
+        <register type="NOA" address="0x00009888" value="0x0CB95000" />
+        <register type="NOA" address="0x00009888" value="0x00B94000" />
+        <register type="NOA" address="0x00009888" value="0x02B91000" />
+        <register type="NOA" address="0x00009888" value="0x0D88C000" />
+        <register type="NOA" address="0x00009888" value="0x0F880003" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A8020" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x238B0520" />
+        <register type="NOA" address="0x00009888" value="0x258BA950" />
+        <register type="NOA" address="0x00009888" value="0x278B0016" />
+        <register type="NOA" address="0x00009888" value="0x198C5400" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0001" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038D2000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA0" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x03835180" />
+        <register type="NOA" address="0x00009888" value="0x05834022" />
+        <register type="NOA" address="0x00009888" value="0x11830000" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x07830000" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x05844000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C137" />
+        <register type="NOA" address="0x00009888" value="0x1D80C147" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x15804000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D801000" />
+        <register type="NOA" address="0x00009888" value="0x4F800111" />
+        <register type="NOA" address="0x00009888" value="0x43800842" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800840" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800800" />
+        <register type="NOA" address="0x00009888" value="0x418014A2" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFFA" />
+        <register type="OA" address="0x00002774" value="0x0000FEFE" />
+        <register type="OA" address="0x00002778" value="0x0007FFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000FEFD" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000FBEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000FBDF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00101100" />
+        <register type="FLEX" address="0x0000E45C" value="0x00201200" />
+        <register type="FLEX" address="0x0000E55C" value="0x00301300" />
+        <register type="FLEX" address="0x0000E65C" value="0x00401400" />
+    </register_config>
+  </set>
+
+  <set name="Data Port Reads Coalescing Gen8"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="data_port_reads_coalescing"
+       hw_config_guid="857fc630-2f09-4804-85f1-084adfadd5ab"
+       chipset="BDW"
+       symbol_name="DataPortReadsCoalescing"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU to Data Port 0 Reads 64"
+             description="The subslice 0 EU data reads from Data Port with 64B per message."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_hdc0_reads64_b"
+             units="messages"
+             symbol_name="EuHdc0Reads64B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Data Port 0 to L3 Data Reads"
+             description="The subslice 0 Data Port data and constant reads from L3 cache."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="hdc0_l3_data_reads"
+             units="messages"
+             symbol_name="Hdc0L3DataReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Data Port 0 to L3 Data Writes"
+             description="The subslice 0 Data Port data writes to L3 cache."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="hdc0_l3_data_writes"
+             units="messages"
+             symbol_name="Hdc0L3DataWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU to Data Port 0 Reads 128"
+             description="The subslice 0 EU data reads from Data Port with 128B per message."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_hdc0_reads128_b"
+             units="messages"
+             symbol_name="EuHdc0Reads128B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="All Data Port 0 Writes to L3"
+             description="The subslice 0 Data Port writes to L3 cache."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="hdc0_l3_writes"
+             units="messages"
+             symbol_name="Hdc0L3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU to Data Port 0 Reads 32"
+             description="The subslice 0 EU data reads from Data Port with 32B per message."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_hdc0_reads32_b"
+             units="messages"
+             symbol_name="EuHdc0Reads32B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU to Data Port 0 Reads 256"
+             description="The subslice 0 EU data reads from Data Port with 256B per message."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_hdc0_reads256_b"
+             units="messages"
+             symbol_name="EuHdc0Reads256B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuBytesReadPerCacheLine"
+             description="Average EU bytes read per L3 cache line."
+             data_type="float"
+             equation="$EuHdc0Reads32B 32 UMUL $EuHdc0Reads64B 64 UMUL $EuHdc0Reads128B 128 UMUL $EuHdc0Reads256B 256 UMUL UADD UADD UADD $Hdc0L3DataReads FDIV"
+             underscore_name="eu_bytes_read_per_cache_line"
+             units="eu bytes per l3 cache line"
+             symbol_name="EuBytesReadPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuDataReadsPerCacheLine"
+             description="Coalescing ratio of EU read requests to L3 cache lines."
+             data_type="float"
+             equation="$EuBytesReadPerCacheLine 64 FDIV"
+             underscore_name="eu_data_reads_per_cache_line"
+             units="utilization"
+             symbol_name="EuDataReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="All Data Port 0 Reads from L3"
+             description="The subslice 0 Data Port reads from L3 cache."
+             data_type="uint64"
+             equation="C 3 READ C 2 READ USUB"
+             underscore_name="hdc0_l3_reads"
+             units="messages"
+             symbol_name="Hdc0L3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x01 AND"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x103D0005" />
+        <register type="NOA" address="0x00009888" value="0x163D240B" />
+        <register type="NOA" address="0x00009888" value="0x1058022F" />
+        <register type="NOA" address="0x00009888" value="0x185B5520" />
+        <register type="NOA" address="0x00009888" value="0x198B0003" />
+        <register type="NOA" address="0x00009888" value="0x005CC000" />
+        <register type="NOA" address="0x00009888" value="0x065CC000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0E5CC000" />
+        <register type="NOA" address="0x00009888" value="0x025C4000" />
+        <register type="NOA" address="0x00009888" value="0x045C8000" />
+        <register type="NOA" address="0x00009888" value="0x003D0000" />
+        <register type="NOA" address="0x00009888" value="0x063D00B0" />
+        <register type="NOA" address="0x00009888" value="0x083D0182" />
+        <register type="NOA" address="0x00009888" value="0x0A3D10A0" />
+        <register type="NOA" address="0x00009888" value="0x0C3D11A2" />
+        <register type="NOA" address="0x00009888" value="0x0E3D0000" />
+        <register type="NOA" address="0x00009888" value="0x183D0000" />
+        <register type="NOA" address="0x00009888" value="0x1A3D0000" />
+        <register type="NOA" address="0x00009888" value="0x0E582242" />
+        <register type="NOA" address="0x00009888" value="0x00586700" />
+        <register type="NOA" address="0x00009888" value="0x0258004F" />
+        <register type="NOA" address="0x00009888" value="0x0658C000" />
+        <register type="NOA" address="0x00009888" value="0x0858C000" />
+        <register type="NOA" address="0x00009888" value="0x0A58C000" />
+        <register type="NOA" address="0x00009888" value="0x0C58C000" />
+        <register type="NOA" address="0x00009888" value="0x045B6300" />
+        <register type="NOA" address="0x00009888" value="0x105B0000" />
+        <register type="NOA" address="0x00009888" value="0x005B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5B0155" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B0000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAAA0" />
+        <register type="NOA" address="0x00009888" value="0x101F02AA" />
+        <register type="NOA" address="0x00009888" value="0x00384000" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18381555" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x04384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0C384000" />
+        <register type="NOA" address="0x00009888" value="0x0039A000" />
+        <register type="NOA" address="0x00009888" value="0x0639A000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x02392000" />
+        <register type="NOA" address="0x00009888" value="0x04398000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x038B6300" />
+        <register type="NOA" address="0x00009888" value="0x058B0062" />
+        <register type="NOA" address="0x00009888" value="0x118B0000" />
+        <register type="NOA" address="0x00009888" value="0x238B02A0" />
+        <register type="NOA" address="0x00009888" value="0x258B5555" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D801000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800001" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800041" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0xBA98BA98" />
+        <register type="OA" address="0x00002748" value="0xBA98BA98" />
+        <register type="OA" address="0x00002744" value="0x00003377" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFF2" />
+        <register type="OA" address="0x00002774" value="0x00007FF0" />
+        <register type="OA" address="0x00002778" value="0x0007FFE2" />
+        <register type="OA" address="0x0000277C" value="0x00007FF0" />
+        <register type="OA" address="0x00002780" value="0x0007FFC2" />
+        <register type="OA" address="0x00002784" value="0x00007FF0" />
+        <register type="OA" address="0x00002788" value="0x0007FF82" />
+        <register type="OA" address="0x0000278C" value="0x00007FF0" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000BFEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000BFDF" />
+        <register type="OA" address="0x000027A0" value="0x0007FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000BFBF" />
+        <register type="OA" address="0x000027A8" value="0x0007FFFA" />
+        <register type="OA" address="0x000027AC" value="0x0000BF7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Data Port Writes Coalescing Gen8"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="data_port_writes_coalescing"
+       hw_config_guid="343ebc99-4a55-414c-8c17-d8e259cf5e20"
+       chipset="BDW"
+       symbol_name="DataPortWritesCoalescing"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU to Data Port 0 Writes 128"
+             description="The subslice 0 EU data simd16 writes to Data Port with 192B per message."
+             data_type="uint64"
+             equation="B 6 READ 2 UDIV"
+             underscore_name="eu_hdc0_writes192_b"
+             units="messages"
+             symbol_name="EuHdc0Writes192B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="EU to Data Port 0 Writes 32B"
+             description="The subslice 0 EU data writes to Data Port with 32B per message."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_hdc0_writes32_b"
+             units="messages"
+             symbol_name="EuHdc0Writes32B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU to Data Port 0 Writes 256B"
+             description="The subslice 0 EU data simd16 writes to Data Port with 256B per message."
+             data_type="uint64"
+             equation="B 7 READ 2 UDIV"
+             underscore_name="eu_hdc0_writes256_b_simd16"
+             units="messages"
+             symbol_name="EuHdc0Writes256BSimd16"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Data Port 0 to L3 Data Reads"
+             description="The subslice 0 Data Port data and constant reads from L3 cache."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="hdc0_l3_data_reads"
+             units="messages"
+             symbol_name="Hdc0L3DataReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Data Port 0 to L3 Data Writes"
+             description="The subslice 0 Data Port data writes to L3 cache."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="hdc0_l3_data_writes"
+             units="messages"
+             symbol_name="Hdc0L3DataWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU to Data Port 0 Writes 64B"
+             description="The subslice 0 EU data simd16 writes to Data Port with 128B per message."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_hdc0_writes128_b_simd16"
+             units="messages"
+             symbol_name="EuHdc0Writes128BSimd16"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU to Data Port 0 Writes 64B"
+             description="The subslice 0 EU data writes to Data Port with 64B per message."
+             data_type="uint64"
+             equation="B 2 READ 2 UDIV"
+             underscore_name="eu_hdc0_writes96_b"
+             units="messages"
+             symbol_name="EuHdc0Writes96B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="All Data Port 0 Writes to L3"
+             description="The subslice 0 Data Port writes to L3 cache."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="hdc0_l3_writes"
+             units="messages"
+             symbol_name="Hdc0L3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU to Data Port 0 Writes 64B"
+             description="The subslice 0 EU data writes to Data Port with 64B per message."
+             data_type="uint64"
+             equation="B 1 READ B 4 READ UADD"
+             underscore_name="eu_hdc0_writes64_b"
+             units="messages"
+             symbol_name="EuHdc0Writes64B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU to Data Port 0 Writes 128"
+             description="The subslice 0 EU data writes to Data Port with 128B per message."
+             data_type="uint64"
+             equation="B 3 READ 2 UDIV"
+             underscore_name="eu_hdc0_writes128_b"
+             units="messages"
+             symbol_name="EuHdc0Writes128B"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuBytesWrittenPerCacheLine"
+             description="Average EU bytes written per L3 cache line."
+             data_type="float"
+             equation="$EuHdc0Writes32B 32 UMUL $EuHdc0Writes64B 64 UMUL $EuHdc0Writes96B 96 UMUL $EuHdc0Writes128B 128 UMUL $EuHdc0Writes128BSimd16 128 UMUL $EuHdc0Writes256BSimd16 256 UMUL $EuHdc0Writes192B 192 UMUL UADD UADD UADD UADD UADD UADD $Hdc0L3DataWrites FDIV"
+             underscore_name="eu_bytes_written_per_cache_line"
+             units="eu bytes per l3 cache line"
+             symbol_name="EuBytesWrittenPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuDataWritesPerCacheLine"
+             description="Coalescing ratio of EU write requests to L3 cache lines."
+             data_type="float"
+             equation="$EuBytesWrittenPerCacheLine 64 FDIV"
+             underscore_name="eu_data_writes_per_cache_line"
+             units="utilization"
+             symbol_name="EuDataWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL 2 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="All Data Port 0 Reads from L3"
+             description="The subslice 0 Data Port reads from L3 cache."
+             data_type="uint64"
+             equation="C 3 READ C 2 READ USUB"
+             underscore_name="hdc0_l3_reads"
+             units="messages"
+             symbol_name="Hdc0L3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x01 AND"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x103D0005" />
+        <register type="NOA" address="0x00009888" value="0x143D0120" />
+        <register type="NOA" address="0x00009888" value="0x163D2400" />
+        <register type="NOA" address="0x00009888" value="0x1058022F" />
+        <register type="NOA" address="0x00009888" value="0x105B0000" />
+        <register type="NOA" address="0x00009888" value="0x198B0003" />
+        <register type="NOA" address="0x00009888" value="0x005CC000" />
+        <register type="NOA" address="0x00009888" value="0x065CC000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0E5CC000" />
+        <register type="NOA" address="0x00009888" value="0x025C4000" />
+        <register type="NOA" address="0x00009888" value="0x045C8000" />
+        <register type="NOA" address="0x00009888" value="0x003D0000" />
+        <register type="NOA" address="0x00009888" value="0x063D0094" />
+        <register type="NOA" address="0x00009888" value="0x083D0182" />
+        <register type="NOA" address="0x00009888" value="0x0A3D1814" />
+        <register type="NOA" address="0x00009888" value="0x0E3D0000" />
+        <register type="NOA" address="0x00009888" value="0x183D0000" />
+        <register type="NOA" address="0x00009888" value="0x1A3D0000" />
+        <register type="NOA" address="0x00009888" value="0x0C3D0000" />
+        <register type="NOA" address="0x00009888" value="0x0E582242" />
+        <register type="NOA" address="0x00009888" value="0x00586700" />
+        <register type="NOA" address="0x00009888" value="0x0258004F" />
+        <register type="NOA" address="0x00009888" value="0x0658C000" />
+        <register type="NOA" address="0x00009888" value="0x0858C000" />
+        <register type="NOA" address="0x00009888" value="0x0A58C000" />
+        <register type="NOA" address="0x00009888" value="0x045B6A80" />
+        <register type="NOA" address="0x00009888" value="0x005B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x185B5400" />
+        <register type="NOA" address="0x00009888" value="0x1A5B0141" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B0000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAAA0" />
+        <register type="NOA" address="0x00009888" value="0x101F0282" />
+        <register type="NOA" address="0x00009888" value="0x00384000" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18381415" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x04384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0C384000" />
+        <register type="NOA" address="0x00009888" value="0x0039A000" />
+        <register type="NOA" address="0x00009888" value="0x0639A000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x02392000" />
+        <register type="NOA" address="0x00009888" value="0x04398000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A82A0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x038B6300" />
+        <register type="NOA" address="0x00009888" value="0x058B0062" />
+        <register type="NOA" address="0x00009888" value="0x118B0000" />
+        <register type="NOA" address="0x00009888" value="0x238B02A0" />
+        <register type="NOA" address="0x00009888" value="0x258B1555" />
+        <register type="NOA" address="0x00009888" value="0x278B0014" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x21852AAA" />
+        <register type="NOA" address="0x00009888" value="0x23850028" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830141" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00000D24" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x4D801000" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800001" />
+        <register type="NOA" address="0x00009888" value="0x43800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800420" />
+        <register type="NOA" address="0x00009888" value="0x3F800421" />
+        <register type="NOA" address="0x00009888" value="0x41800041" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0xBA98BA98" />
+        <register type="OA" address="0x00002748" value="0xBA98BA98" />
+        <register type="OA" address="0x00002744" value="0x00003377" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FF72" />
+        <register type="OA" address="0x00002774" value="0x0000BFD0" />
+        <register type="OA" address="0x00002778" value="0x0007FF62" />
+        <register type="OA" address="0x0000277C" value="0x0000BFD0" />
+        <register type="OA" address="0x00002780" value="0x0007FF42" />
+        <register type="OA" address="0x00002784" value="0x0000BFD0" />
+        <register type="OA" address="0x00002788" value="0x0007FF02" />
+        <register type="OA" address="0x0000278C" value="0x0000BFD0" />
+        <register type="OA" address="0x00002790" value="0x0005FFF2" />
+        <register type="OA" address="0x00002794" value="0x0000BFD0" />
+        <register type="OA" address="0x00002798" value="0x0005FFE2" />
+        <register type="OA" address="0x0000279C" value="0x0000BFD0" />
+        <register type="OA" address="0x000027A0" value="0x0005FFC2" />
+        <register type="OA" address="0x000027A4" value="0x0000BFD0" />
+        <register type="OA" address="0x000027A8" value="0x0005FF82" />
+        <register type="OA" address="0x000027AC" value="0x0000BFD0" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="7bdafd88-a4fa-4ed5-bc09-1a977aa5be3e"
+       chipset="BDW"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 1 READ C 0 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader12_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader12AccessStalledOnL3"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader02_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader02AccessStalledOnL3"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ B 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader10_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader10AccessStalledOnL3"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="B 7 READ B 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader11_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader11AccessStalledOnL3"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ C 2 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x105C0232" />
+        <register type="NOA" address="0x00009888" value="0x10580232" />
+        <register type="NOA" address="0x00009888" value="0x10380232" />
+        <register type="NOA" address="0x00009888" value="0x10DC0232" />
+        <register type="NOA" address="0x00009888" value="0x10D80232" />
+        <register type="NOA" address="0x00009888" value="0x10B80232" />
+        <register type="NOA" address="0x00009888" value="0x118E4400" />
+        <register type="NOA" address="0x00009888" value="0x025C6080" />
+        <register type="NOA" address="0x00009888" value="0x045C004B" />
+        <register type="NOA" address="0x00009888" value="0x005C8000" />
+        <register type="NOA" address="0x00009888" value="0x00582080" />
+        <register type="NOA" address="0x00009888" value="0x0258004B" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00AA" />
+        <register type="NOA" address="0x00009888" value="0x04386080" />
+        <register type="NOA" address="0x00009888" value="0x0638404B" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A380000" />
+        <register type="NOA" address="0x00009888" value="0x0C380000" />
+        <register type="NOA" address="0x00009888" value="0x00398000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x0CDC25C1" />
+        <register type="NOA" address="0x00009888" value="0x0ADCC000" />
+        <register type="NOA" address="0x00009888" value="0x0AD825C1" />
+        <register type="NOA" address="0x00009888" value="0x18DB4000" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0001" />
+        <register type="NOA" address="0x00009888" value="0x0E9F8000" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0EB825C1" />
+        <register type="NOA" address="0x00009888" value="0x18B80154" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0D88C000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258BAA05" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x198C5400" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x098DC000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x098E05C0" />
+        <register type="NOA" address="0x00009888" value="0x058E0000" />
+        <register type="NOA" address="0x00009888" value="0x198F0020" />
+        <register type="NOA" address="0x00009888" value="0x2185AA0A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x19835000" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x19808000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x51800040" />
+        <register type="NOA" address="0x00009888" value="0x43800400" />
+        <register type="NOA" address="0x00009888" value="0x45800800" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800C62" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F801042" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x418014A4" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FFF7" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="9385ebb2-f34f-4aa5-aec5-7e9cbbea0f0b"
+       chipset="BDW"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank1 Stalled"
+             description="The percentage of time in which slice1 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank1_stalled"
+             units="percent"
+             symbol_name="L31Bank1Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice1 L3 Bank0 Stalled"
+             description="The percentage of time in which slice1 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank0_stalled"
+             units="percent"
+             symbol_name="L31Bank0Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Slice1 L3 Bank1 Active"
+             description="The percentage of time in which slice1 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank1_active"
+             units="percent"
+             symbol_name="L31Bank1Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank0 Active"
+             description="The percentage of time in which slice1 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank0_active"
+             units="percent"
+             symbol_name="L31Bank0Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x10BF03DA" />
+        <register type="NOA" address="0x00009888" value="0x14BF0001" />
+        <register type="NOA" address="0x00009888" value="0x12980340" />
+        <register type="NOA" address="0x00009888" value="0x12990340" />
+        <register type="NOA" address="0x00009888" value="0x0CBF1187" />
+        <register type="NOA" address="0x00009888" value="0x0EBF1205" />
+        <register type="NOA" address="0x00009888" value="0x00BF0500" />
+        <register type="NOA" address="0x00009888" value="0x02BF042B" />
+        <register type="NOA" address="0x00009888" value="0x04BF002C" />
+        <register type="NOA" address="0x00009888" value="0x0CDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDAC000" />
+        <register type="NOA" address="0x00009888" value="0x00DA8000" />
+        <register type="NOA" address="0x00009888" value="0x02DAC000" />
+        <register type="NOA" address="0x00009888" value="0x04DA4000" />
+        <register type="NOA" address="0x00009888" value="0x04983400" />
+        <register type="NOA" address="0x00009888" value="0x10980000" />
+        <register type="NOA" address="0x00009888" value="0x06990034" />
+        <register type="NOA" address="0x00009888" value="0x10990000" />
+        <register type="NOA" address="0x00009888" value="0x0C9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E9DC000" />
+        <register type="NOA" address="0x00009888" value="0x009D8000" />
+        <register type="NOA" address="0x00009888" value="0x029DC000" />
+        <register type="NOA" address="0x00009888" value="0x049D4000" />
+        <register type="NOA" address="0x00009888" value="0x109F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F00BA" />
+        <register type="NOA" address="0x00009888" value="0x0CB88000" />
+        <register type="NOA" address="0x00009888" value="0x0CB95000" />
+        <register type="NOA" address="0x00009888" value="0x0EB95000" />
+        <register type="NOA" address="0x00009888" value="0x00B94000" />
+        <register type="NOA" address="0x00009888" value="0x02B95000" />
+        <register type="NOA" address="0x00009888" value="0x04B91000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x0CBA4000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x258B800A" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B5500" />
+        <register type="NOA" address="0x00009888" value="0x198C4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x47800000" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_2"
+       hw_config_guid="446ae59b-ff2e-41c9-b49e-0184a54bf00a"
+       chipset="BDW"
+       symbol_name="L3_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x103F03DA" />
+        <register type="NOA" address="0x00009888" value="0x143F0001" />
+        <register type="NOA" address="0x00009888" value="0x12180340" />
+        <register type="NOA" address="0x00009888" value="0x12190340" />
+        <register type="NOA" address="0x00009888" value="0x0C3F1187" />
+        <register type="NOA" address="0x00009888" value="0x0E3F1205" />
+        <register type="NOA" address="0x00009888" value="0x003F0500" />
+        <register type="NOA" address="0x00009888" value="0x023F042B" />
+        <register type="NOA" address="0x00009888" value="0x043F002C" />
+        <register type="NOA" address="0x00009888" value="0x0C5AC000" />
+        <register type="NOA" address="0x00009888" value="0x0E5AC000" />
+        <register type="NOA" address="0x00009888" value="0x005A8000" />
+        <register type="NOA" address="0x00009888" value="0x025AC000" />
+        <register type="NOA" address="0x00009888" value="0x045A4000" />
+        <register type="NOA" address="0x00009888" value="0x04183400" />
+        <register type="NOA" address="0x00009888" value="0x10180000" />
+        <register type="NOA" address="0x00009888" value="0x06190034" />
+        <register type="NOA" address="0x00009888" value="0x10190000" />
+        <register type="NOA" address="0x00009888" value="0x0C1DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1DC000" />
+        <register type="NOA" address="0x00009888" value="0x001D8000" />
+        <register type="NOA" address="0x00009888" value="0x021DC000" />
+        <register type="NOA" address="0x00009888" value="0x041D4000" />
+        <register type="NOA" address="0x00009888" value="0x101F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00BA" />
+        <register type="NOA" address="0x00009888" value="0x0C388000" />
+        <register type="NOA" address="0x00009888" value="0x0C395000" />
+        <register type="NOA" address="0x00009888" value="0x0E395000" />
+        <register type="NOA" address="0x00009888" value="0x00394000" />
+        <register type="NOA" address="0x00009888" value="0x02395000" />
+        <register type="NOA" address="0x00009888" value="0x04391000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AA800" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258B4005" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x47800000" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_3"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_3"
+       hw_config_guid="84a7956f-1ea4-4d0d-837f-e39a0376e38c"
+       chipset="BDW"
+       symbol_name="L3_3"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 L3 Bank3 Stalled"
+             description="The percentage of time in which slice0 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_stalled"
+             units="percent"
+             symbol_name="L30Bank3Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank3 Active"
+             description="The percentage of time in which slice0 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_active"
+             units="percent"
+             symbol_name="L30Bank3Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank3 Active"
+             description="The percentage of time in which slice1 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank3_active"
+             units="percent"
+             symbol_name="L31Bank3Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice1 L3 Bank3 Stalled"
+             description="The percentage of time in which slice1 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank3_stalled"
+             units="percent"
+             symbol_name="L31Bank3Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121B0340" />
+        <register type="NOA" address="0x00009888" value="0x103F0274" />
+        <register type="NOA" address="0x00009888" value="0x123F0000" />
+        <register type="NOA" address="0x00009888" value="0x129B0340" />
+        <register type="NOA" address="0x00009888" value="0x10BF0274" />
+        <register type="NOA" address="0x00009888" value="0x12BF0000" />
+        <register type="NOA" address="0x00009888" value="0x041B3400" />
+        <register type="NOA" address="0x00009888" value="0x101B0000" />
+        <register type="NOA" address="0x00009888" value="0x045C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A3D4000" />
+        <register type="NOA" address="0x00009888" value="0x003F0080" />
+        <register type="NOA" address="0x00009888" value="0x023F0793" />
+        <register type="NOA" address="0x00009888" value="0x043F0014" />
+        <register type="NOA" address="0x00009888" value="0x04588000" />
+        <register type="NOA" address="0x00009888" value="0x005A8000" />
+        <register type="NOA" address="0x00009888" value="0x025AC000" />
+        <register type="NOA" address="0x00009888" value="0x045A4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x001D8000" />
+        <register type="NOA" address="0x00009888" value="0x021DC000" />
+        <register type="NOA" address="0x00009888" value="0x041D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F002A" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x00394000" />
+        <register type="NOA" address="0x00009888" value="0x02395000" />
+        <register type="NOA" address="0x00009888" value="0x04399000" />
+        <register type="NOA" address="0x00009888" value="0x069B0034" />
+        <register type="NOA" address="0x00009888" value="0x109B0000" />
+        <register type="NOA" address="0x00009888" value="0x06DC4000" />
+        <register type="NOA" address="0x00009888" value="0x0CBD4000" />
+        <register type="NOA" address="0x00009888" value="0x0CBF0981" />
+        <register type="NOA" address="0x00009888" value="0x0EBF0A0F" />
+        <register type="NOA" address="0x00009888" value="0x06D84000" />
+        <register type="NOA" address="0x00009888" value="0x0CDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0CDB4000" />
+        <register type="NOA" address="0x00009888" value="0x0C9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E9DC000" />
+        <register type="NOA" address="0x00009888" value="0x109F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0E9F0080" />
+        <register type="NOA" address="0x00009888" value="0x0CB84000" />
+        <register type="NOA" address="0x00009888" value="0x0CB95000" />
+        <register type="NOA" address="0x00009888" value="0x0EB95000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258B8009" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x198C4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800C00" />
+        <register type="NOA" address="0x00009888" value="0x47800C63" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F8014A5" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800045" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_4"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_4"
+       hw_config_guid="92b493d9-df18-4bed-be06-5cac6f2a6f5f"
+       chipset="BDW"
+       symbol_name="L3_4"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Stalled"
+             description="The percentage of time in which slice0 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_stalled"
+             units="percent"
+             symbol_name="L30Bank2Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank2 Active"
+             description="The percentage of time in which slice1 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank2_active"
+             units="percent"
+             symbol_name="L31Bank2Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 L3 Bank2 Active"
+             description="The percentage of time in which slice0 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_active"
+             units="percent"
+             symbol_name="L30Bank2Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Slice1 L3 Bank2 Stalled"
+             description="The percentage of time in which slice1 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank2_stalled"
+             units="percent"
+             symbol_name="L31Bank2Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121A0340" />
+        <register type="NOA" address="0x00009888" value="0x103F0017" />
+        <register type="NOA" address="0x00009888" value="0x123F0020" />
+        <register type="NOA" address="0x00009888" value="0x129A0340" />
+        <register type="NOA" address="0x00009888" value="0x10BF0017" />
+        <register type="NOA" address="0x00009888" value="0x12BF0020" />
+        <register type="NOA" address="0x00009888" value="0x041A3400" />
+        <register type="NOA" address="0x00009888" value="0x101A0000" />
+        <register type="NOA" address="0x00009888" value="0x043B8000" />
+        <register type="NOA" address="0x00009888" value="0x0A3E0010" />
+        <register type="NOA" address="0x00009888" value="0x003F0200" />
+        <register type="NOA" address="0x00009888" value="0x023F0113" />
+        <register type="NOA" address="0x00009888" value="0x043F0014" />
+        <register type="NOA" address="0x00009888" value="0x02592000" />
+        <register type="NOA" address="0x00009888" value="0x005A8000" />
+        <register type="NOA" address="0x00009888" value="0x025AC000" />
+        <register type="NOA" address="0x00009888" value="0x045A4000" />
+        <register type="NOA" address="0x00009888" value="0x0A1C8000" />
+        <register type="NOA" address="0x00009888" value="0x001D8000" />
+        <register type="NOA" address="0x00009888" value="0x021DC000" />
+        <register type="NOA" address="0x00009888" value="0x041D4000" />
+        <register type="NOA" address="0x00009888" value="0x0A1E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F001A" />
+        <register type="NOA" address="0x00009888" value="0x00394000" />
+        <register type="NOA" address="0x00009888" value="0x02395000" />
+        <register type="NOA" address="0x00009888" value="0x04391000" />
+        <register type="NOA" address="0x00009888" value="0x069A0034" />
+        <register type="NOA" address="0x00009888" value="0x109A0000" />
+        <register type="NOA" address="0x00009888" value="0x06BB4000" />
+        <register type="NOA" address="0x00009888" value="0x0ABE0040" />
+        <register type="NOA" address="0x00009888" value="0x0CBF0984" />
+        <register type="NOA" address="0x00009888" value="0x0EBF0A02" />
+        <register type="NOA" address="0x00009888" value="0x02D94000" />
+        <register type="NOA" address="0x00009888" value="0x0CDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0C9C0400" />
+        <register type="NOA" address="0x00009888" value="0x0C9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0C9E0400" />
+        <register type="NOA" address="0x00009888" value="0x109F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0E9F0040" />
+        <register type="NOA" address="0x00009888" value="0x0CB95000" />
+        <register type="NOA" address="0x00009888" value="0x0EB95000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258B8009" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x198C4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800800" />
+        <register type="NOA" address="0x00009888" value="0x47800842" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F801084" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800044" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="14345c35-cc46-40d0-bb04-6ed1fbb43679"
+       chipset="BDW"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 Rasterizer Input Available"
+             description="The percentage of time in which slice1 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer1_input_available"
+             units="percent"
+             symbol_name="Rasterizer1InputAvailable"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 Pixel Values Ready"
+             description="The percentage of time in which slice1 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values1_ready"
+             units="percent"
+             symbol_name="PixelValues1Ready"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice1 PS Output Available"
+             description="The percentage of time in which slice1 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output1_available"
+             units="percent"
+             symbol_name="PSOutput1Available"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice1 Rasterizer Output Ready"
+             description="The percentage of time in which slice1 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer1_output_ready"
+             units="percent"
+             symbol_name="Rasterizer1OutputReady"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice1 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice1 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data1_ready"
+             units="percent"
+             symbol_name="PixelData1Ready"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x143B000E" />
+        <register type="NOA" address="0x00009888" value="0x043C55C0" />
+        <register type="NOA" address="0x00009888" value="0x0A1E0280" />
+        <register type="NOA" address="0x00009888" value="0x0C1E0408" />
+        <register type="NOA" address="0x00009888" value="0x10390000" />
+        <register type="NOA" address="0x00009888" value="0x12397A1F" />
+        <register type="NOA" address="0x00009888" value="0x14BB000E" />
+        <register type="NOA" address="0x00009888" value="0x04BC5000" />
+        <register type="NOA" address="0x00009888" value="0x0A9E0296" />
+        <register type="NOA" address="0x00009888" value="0x0C9E0008" />
+        <register type="NOA" address="0x00009888" value="0x10B90000" />
+        <register type="NOA" address="0x00009888" value="0x12B97A1F" />
+        <register type="NOA" address="0x00009888" value="0x063B0042" />
+        <register type="NOA" address="0x00009888" value="0x103B0000" />
+        <register type="NOA" address="0x00009888" value="0x083C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A3E0040" />
+        <register type="NOA" address="0x00009888" value="0x043F8000" />
+        <register type="NOA" address="0x00009888" value="0x02594000" />
+        <register type="NOA" address="0x00009888" value="0x045A8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1C0400" />
+        <register type="NOA" address="0x00009888" value="0x041D8000" />
+        <register type="NOA" address="0x00009888" value="0x081E02C0" />
+        <register type="NOA" address="0x00009888" value="0x0E1E0000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1F0260" />
+        <register type="NOA" address="0x00009888" value="0x101F0014" />
+        <register type="NOA" address="0x00009888" value="0x003905E0" />
+        <register type="NOA" address="0x00009888" value="0x06390BC0" />
+        <register type="NOA" address="0x00009888" value="0x02390018" />
+        <register type="NOA" address="0x00009888" value="0x04394000" />
+        <register type="NOA" address="0x00009888" value="0x04BB0042" />
+        <register type="NOA" address="0x00009888" value="0x10BB0000" />
+        <register type="NOA" address="0x00009888" value="0x02BC05C0" />
+        <register type="NOA" address="0x00009888" value="0x08BC0000" />
+        <register type="NOA" address="0x00009888" value="0x0ABE0004" />
+        <register type="NOA" address="0x00009888" value="0x02BF8000" />
+        <register type="NOA" address="0x00009888" value="0x02D91000" />
+        <register type="NOA" address="0x00009888" value="0x02DA8000" />
+        <register type="NOA" address="0x00009888" value="0x089C8000" />
+        <register type="NOA" address="0x00009888" value="0x029D8000" />
+        <register type="NOA" address="0x00009888" value="0x089E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E9E0000" />
+        <register type="NOA" address="0x00009888" value="0x0E9FA806" />
+        <register type="NOA" address="0x00009888" value="0x109F0142" />
+        <register type="NOA" address="0x00009888" value="0x08B90617" />
+        <register type="NOA" address="0x00009888" value="0x0AB90BE0" />
+        <register type="NOA" address="0x00009888" value="0x02B94000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000C" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A2800" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B52A0" />
+        <register type="NOA" address="0x00009888" value="0x258B6A95" />
+        <register type="NOA" address="0x00009888" value="0x278B0029" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C1500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0014" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x038D8000" />
+        <register type="NOA" address="0x00009888" value="0x058D2000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x4D800444" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F804000" />
+        <register type="NOA" address="0x00009888" value="0x43801080" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800084" />
+        <register type="NOA" address="0x00009888" value="0x53800044" />
+        <register type="NOA" address="0x00009888" value="0x47801080" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800840" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00006000" />
+        <register type="OA" address="0x00002774" value="0x0000F3FF" />
+        <register type="OA" address="0x00002778" value="0x00001800" />
+        <register type="OA" address="0x0000277C" value="0x0000FCFF" />
+        <register type="OA" address="0x00002780" value="0x00000600" />
+        <register type="OA" address="0x00002784" value="0x0000FF3F" />
+        <register type="OA" address="0x00002788" value="0x00000180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000060" />
+        <register type="OA" address="0x00002794" value="0x0000FFF3" />
+        <register type="OA" address="0x00002798" value="0x00000018" />
+        <register type="OA" address="0x0000279C" value="0x0000FFFC" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler_1"
+       hw_config_guid="f0c6ba37-d3d3-4211-91b5-226730312a54"
+       chipset="BDW"
+       symbol_name="Sampler_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Slice1 Subslice0 Input Available"
+             description="The percentage of time in which slice1 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler10_input_available"
+             units="percent"
+             symbol_name="Sampler10InputAvailable"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Slice1 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice1 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler12_output_ready"
+             units="percent"
+             symbol_name="Sampler12OutputReady"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 Subslice1 Input Available"
+             description="The percentage of time in which slice1 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler11_input_available"
+             units="percent"
+             symbol_name="Sampler11InputAvailable"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice1 Subslice2 Input Available"
+             description="The percentage of time in which slice1 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler12_input_available"
+             units="percent"
+             symbol_name="Sampler12InputAvailable"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice1 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice1 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler10_output_ready"
+             units="percent"
+             symbol_name="Sampler10OutputReady"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Slice1 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice1 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler11_output_ready"
+             units="percent"
+             symbol_name="Sampler11OutputReady"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x18921400" />
+        <register type="NOA" address="0x00009888" value="0x149500AB" />
+        <register type="NOA" address="0x00009888" value="0x18B21400" />
+        <register type="NOA" address="0x00009888" value="0x14B500AB" />
+        <register type="NOA" address="0x00009888" value="0x18D21400" />
+        <register type="NOA" address="0x00009888" value="0x14D500AB" />
+        <register type="NOA" address="0x00009888" value="0x0CDC8000" />
+        <register type="NOA" address="0x00009888" value="0x0EDC4000" />
+        <register type="NOA" address="0x00009888" value="0x02DCC000" />
+        <register type="NOA" address="0x00009888" value="0x04DCC000" />
+        <register type="NOA" address="0x00009888" value="0x1ABD00A0" />
+        <register type="NOA" address="0x00009888" value="0x0ABD8000" />
+        <register type="NOA" address="0x00009888" value="0x0CD88000" />
+        <register type="NOA" address="0x00009888" value="0x0ED84000" />
+        <register type="NOA" address="0x00009888" value="0x04D88000" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0050" />
+        <register type="NOA" address="0x00009888" value="0x04DB8000" />
+        <register type="NOA" address="0x00009888" value="0x06DB8000" />
+        <register type="NOA" address="0x00009888" value="0x08DB8000" />
+        <register type="NOA" address="0x00009888" value="0x0ADB4000" />
+        <register type="NOA" address="0x00009888" value="0x109F02A0" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F00AA" />
+        <register type="NOA" address="0x00009888" value="0x18B82500" />
+        <register type="NOA" address="0x00009888" value="0x02B88000" />
+        <register type="NOA" address="0x00009888" value="0x04B84000" />
+        <register type="NOA" address="0x00009888" value="0x06B84000" />
+        <register type="NOA" address="0x00009888" value="0x08B84000" />
+        <register type="NOA" address="0x00009888" value="0x0AB84000" />
+        <register type="NOA" address="0x00009888" value="0x0CB88000" />
+        <register type="NOA" address="0x00009888" value="0x0CB98000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x00B98000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x04B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x1ABA0200" />
+        <register type="NOA" address="0x00009888" value="0x02BA8000" />
+        <register type="NOA" address="0x00009888" value="0x0CBA8000" />
+        <register type="NOA" address="0x00009888" value="0x04908000" />
+        <register type="NOA" address="0x00009888" value="0x04918000" />
+        <register type="NOA" address="0x00009888" value="0x04927300" />
+        <register type="NOA" address="0x00009888" value="0x10920000" />
+        <register type="NOA" address="0x00009888" value="0x1893000A" />
+        <register type="NOA" address="0x00009888" value="0x0A934000" />
+        <register type="NOA" address="0x00009888" value="0x0A946000" />
+        <register type="NOA" address="0x00009888" value="0x0C959000" />
+        <register type="NOA" address="0x00009888" value="0x0E950098" />
+        <register type="NOA" address="0x00009888" value="0x10950000" />
+        <register type="NOA" address="0x00009888" value="0x04B04000" />
+        <register type="NOA" address="0x00009888" value="0x04B14000" />
+        <register type="NOA" address="0x00009888" value="0x04B20073" />
+        <register type="NOA" address="0x00009888" value="0x10B20000" />
+        <register type="NOA" address="0x00009888" value="0x04B38000" />
+        <register type="NOA" address="0x00009888" value="0x06B38000" />
+        <register type="NOA" address="0x00009888" value="0x08B34000" />
+        <register type="NOA" address="0x00009888" value="0x04B4C000" />
+        <register type="NOA" address="0x00009888" value="0x02B59890" />
+        <register type="NOA" address="0x00009888" value="0x10B50000" />
+        <register type="NOA" address="0x00009888" value="0x06D04000" />
+        <register type="NOA" address="0x00009888" value="0x06D14000" />
+        <register type="NOA" address="0x00009888" value="0x06D20073" />
+        <register type="NOA" address="0x00009888" value="0x10D20000" />
+        <register type="NOA" address="0x00009888" value="0x18D30020" />
+        <register type="NOA" address="0x00009888" value="0x02D38000" />
+        <register type="NOA" address="0x00009888" value="0x0CD34000" />
+        <register type="NOA" address="0x00009888" value="0x0AD48000" />
+        <register type="NOA" address="0x00009888" value="0x04D42000" />
+        <register type="NOA" address="0x00009888" value="0x0ED59000" />
+        <register type="NOA" address="0x00009888" value="0x00D59800" />
+        <register type="NOA" address="0x00009888" value="0x10D50000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000E" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B5500" />
+        <register type="NOA" address="0x00009888" value="0x258B000A" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8D8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x2185000A" />
+        <register type="NOA" address="0x00009888" value="0x1B830150" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D848000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D808000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47801021" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800C64" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800C02" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler_2"
+       hw_config_guid="30bf3702-48cf-4bca-b412-7cf50bb2f564"
+       chipset="BDW"
+       symbol_name="Sampler_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x18121400" />
+        <register type="NOA" address="0x00009888" value="0x141500AB" />
+        <register type="NOA" address="0x00009888" value="0x18321400" />
+        <register type="NOA" address="0x00009888" value="0x143500AB" />
+        <register type="NOA" address="0x00009888" value="0x18521400" />
+        <register type="NOA" address="0x00009888" value="0x145500AB" />
+        <register type="NOA" address="0x00009888" value="0x0C5C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5C4000" />
+        <register type="NOA" address="0x00009888" value="0x025CC000" />
+        <register type="NOA" address="0x00009888" value="0x045CC000" />
+        <register type="NOA" address="0x00009888" value="0x1A3D00A0" />
+        <register type="NOA" address="0x00009888" value="0x0A3D8000" />
+        <register type="NOA" address="0x00009888" value="0x0C588000" />
+        <register type="NOA" address="0x00009888" value="0x0E584000" />
+        <register type="NOA" address="0x00009888" value="0x04588000" />
+        <register type="NOA" address="0x00009888" value="0x1A5B0050" />
+        <register type="NOA" address="0x00009888" value="0x045B8000" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B8000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x101F02A0" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00AA" />
+        <register type="NOA" address="0x00009888" value="0x18382500" />
+        <register type="NOA" address="0x00009888" value="0x02388000" />
+        <register type="NOA" address="0x00009888" value="0x04384000" />
+        <register type="NOA" address="0x00009888" value="0x06384000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0C388000" />
+        <register type="NOA" address="0x00009888" value="0x0C398000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x00398000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x1A3A0200" />
+        <register type="NOA" address="0x00009888" value="0x023A8000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A8000" />
+        <register type="NOA" address="0x00009888" value="0x04108000" />
+        <register type="NOA" address="0x00009888" value="0x04118000" />
+        <register type="NOA" address="0x00009888" value="0x04127300" />
+        <register type="NOA" address="0x00009888" value="0x10120000" />
+        <register type="NOA" address="0x00009888" value="0x1813000A" />
+        <register type="NOA" address="0x00009888" value="0x0A134000" />
+        <register type="NOA" address="0x00009888" value="0x0A146000" />
+        <register type="NOA" address="0x00009888" value="0x0C159000" />
+        <register type="NOA" address="0x00009888" value="0x0E150098" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x04304000" />
+        <register type="NOA" address="0x00009888" value="0x04314000" />
+        <register type="NOA" address="0x00009888" value="0x04320073" />
+        <register type="NOA" address="0x00009888" value="0x10320000" />
+        <register type="NOA" address="0x00009888" value="0x04338000" />
+        <register type="NOA" address="0x00009888" value="0x06338000" />
+        <register type="NOA" address="0x00009888" value="0x08334000" />
+        <register type="NOA" address="0x00009888" value="0x0434C000" />
+        <register type="NOA" address="0x00009888" value="0x02359890" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x06504000" />
+        <register type="NOA" address="0x00009888" value="0x06514000" />
+        <register type="NOA" address="0x00009888" value="0x06520073" />
+        <register type="NOA" address="0x00009888" value="0x10520000" />
+        <register type="NOA" address="0x00009888" value="0x18530020" />
+        <register type="NOA" address="0x00009888" value="0x02538000" />
+        <register type="NOA" address="0x00009888" value="0x0C534000" />
+        <register type="NOA" address="0x00009888" value="0x0A548000" />
+        <register type="NOA" address="0x00009888" value="0x04542000" />
+        <register type="NOA" address="0x00009888" value="0x0E559000" />
+        <register type="NOA" address="0x00009888" value="0x00559800" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AA000" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x258B0005" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x2185000A" />
+        <register type="NOA" address="0x00009888" value="0x1B830150" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D848000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D808000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47801021" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800C64" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800C02" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="238bec85-df05-44f3-b905-d166712f2451"
+       chipset="BDW"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice1 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice1 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread11_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread11ReadyForDispatch"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice1 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice1 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread12_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread12ReadyForDispatch"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice1 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice1 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread10_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread10ReadyForDispatch"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice1 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice1 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread12_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread12ReadyForDispatch"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice1 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice1 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread10_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread10ReadyForDispatch"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice1 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice1 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread11_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread11ReadyForDispatch"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x16154D60" />
+        <register type="NOA" address="0x00009888" value="0x16352E60" />
+        <register type="NOA" address="0x00009888" value="0x16554D60" />
+        <register type="NOA" address="0x00009888" value="0x16950000" />
+        <register type="NOA" address="0x00009888" value="0x16B50000" />
+        <register type="NOA" address="0x00009888" value="0x16D50000" />
+        <register type="NOA" address="0x00009888" value="0x005C8000" />
+        <register type="NOA" address="0x00009888" value="0x045CC000" />
+        <register type="NOA" address="0x00009888" value="0x065C4000" />
+        <register type="NOA" address="0x00009888" value="0x083D8000" />
+        <register type="NOA" address="0x00009888" value="0x0A3D8000" />
+        <register type="NOA" address="0x00009888" value="0x0458C000" />
+        <register type="NOA" address="0x00009888" value="0x025B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00AA" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x04388000" />
+        <register type="NOA" address="0x00009888" value="0x06388000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0C384000" />
+        <register type="NOA" address="0x00009888" value="0x00398000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x043A8000" />
+        <register type="NOA" address="0x00009888" value="0x063A8000" />
+        <register type="NOA" address="0x00009888" value="0x08138000" />
+        <register type="NOA" address="0x00009888" value="0x0A138000" />
+        <register type="NOA" address="0x00009888" value="0x06143000" />
+        <register type="NOA" address="0x00009888" value="0x0415CFC7" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x02338000" />
+        <register type="NOA" address="0x00009888" value="0x0C338000" />
+        <register type="NOA" address="0x00009888" value="0x04342000" />
+        <register type="NOA" address="0x00009888" value="0x06344000" />
+        <register type="NOA" address="0x00009888" value="0x0035C700" />
+        <register type="NOA" address="0x00009888" value="0x063500CF" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x04538000" />
+        <register type="NOA" address="0x00009888" value="0x06538000" />
+        <register type="NOA" address="0x00009888" value="0x0454C000" />
+        <register type="NOA" address="0x00009888" value="0x0255CFC7" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x06DC8000" />
+        <register type="NOA" address="0x00009888" value="0x08DC4000" />
+        <register type="NOA" address="0x00009888" value="0x0CDCC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDCC000" />
+        <register type="NOA" address="0x00009888" value="0x1ABD00A8" />
+        <register type="NOA" address="0x00009888" value="0x0CD8C000" />
+        <register type="NOA" address="0x00009888" value="0x0ED84000" />
+        <register type="NOA" address="0x00009888" value="0x0EDB8000" />
+        <register type="NOA" address="0x00009888" value="0x18DB0800" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0254" />
+        <register type="NOA" address="0x00009888" value="0x0E9FAA00" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0EB84000" />
+        <register type="NOA" address="0x00009888" value="0x16B84000" />
+        <register type="NOA" address="0x00009888" value="0x18B8156A" />
+        <register type="NOA" address="0x00009888" value="0x06B98000" />
+        <register type="NOA" address="0x00009888" value="0x08B9A000" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x18BAA000" />
+        <register type="NOA" address="0x00009888" value="0x1ABA0002" />
+        <register type="NOA" address="0x00009888" value="0x16934000" />
+        <register type="NOA" address="0x00009888" value="0x1893000A" />
+        <register type="NOA" address="0x00009888" value="0x0A947000" />
+        <register type="NOA" address="0x00009888" value="0x0C95C5C1" />
+        <register type="NOA" address="0x00009888" value="0x0E9500C3" />
+        <register type="NOA" address="0x00009888" value="0x10950000" />
+        <register type="NOA" address="0x00009888" value="0x0EB38000" />
+        <register type="NOA" address="0x00009888" value="0x16B30040" />
+        <register type="NOA" address="0x00009888" value="0x18B30020" />
+        <register type="NOA" address="0x00009888" value="0x06B48000" />
+        <register type="NOA" address="0x00009888" value="0x08B41000" />
+        <register type="NOA" address="0x00009888" value="0x0AB48000" />
+        <register type="NOA" address="0x00009888" value="0x06B5C500" />
+        <register type="NOA" address="0x00009888" value="0x08B500C3" />
+        <register type="NOA" address="0x00009888" value="0x0EB5C100" />
+        <register type="NOA" address="0x00009888" value="0x10B50000" />
+        <register type="NOA" address="0x00009888" value="0x16D31500" />
+        <register type="NOA" address="0x00009888" value="0x08D4E000" />
+        <register type="NOA" address="0x00009888" value="0x08D5C100" />
+        <register type="NOA" address="0x00009888" value="0x0AD5C3C5" />
+        <register type="NOA" address="0x00009888" value="0x10D50000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F800" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258BAAA5" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x0F8C4000" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C5500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x078D8000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800C42" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800063" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800800" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F8014A4" />
+        <register type="NOA" address="0x00009888" value="0x41801042" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x0000FE7F" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFBF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFF7" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FFF9" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="24bf02cd-8693-4583-981c-c4165b33da01"
+       chipset="BDW"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice1 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice1 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header11_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader11ReadyPort1"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice1 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header11_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader11ReadyPort0"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice1 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header12_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader12ReadyPort0"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice1 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header10_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader10ReadyPort1"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice1 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header12_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader12ReadyPort1"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice1 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header10_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader10ReadyPort0"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x16150000" />
+        <register type="NOA" address="0x00009888" value="0x16350000" />
+        <register type="NOA" address="0x00009888" value="0x16550000" />
+        <register type="NOA" address="0x00009888" value="0x16952E60" />
+        <register type="NOA" address="0x00009888" value="0x16B54D60" />
+        <register type="NOA" address="0x00009888" value="0x16D52E60" />
+        <register type="NOA" address="0x00009888" value="0x065C8000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5C4000" />
+        <register type="NOA" address="0x00009888" value="0x0E3D8000" />
+        <register type="NOA" address="0x00009888" value="0x183DA000" />
+        <register type="NOA" address="0x00009888" value="0x06588000" />
+        <register type="NOA" address="0x00009888" value="0x08588000" />
+        <register type="NOA" address="0x00009888" value="0x0A584000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x185B5800" />
+        <register type="NOA" address="0x00009888" value="0x1A5B000A" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA00" />
+        <register type="NOA" address="0x00009888" value="0x101F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18382A55" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x1A3A02A0" />
+        <register type="NOA" address="0x00009888" value="0x0E138000" />
+        <register type="NOA" address="0x00009888" value="0x16130500" />
+        <register type="NOA" address="0x00009888" value="0x06148000" />
+        <register type="NOA" address="0x00009888" value="0x08146000" />
+        <register type="NOA" address="0x00009888" value="0x0615C100" />
+        <register type="NOA" address="0x00009888" value="0x0815C500" />
+        <register type="NOA" address="0x00009888" value="0x0A1500C3" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x16335040" />
+        <register type="NOA" address="0x00009888" value="0x08349000" />
+        <register type="NOA" address="0x00009888" value="0x0A341000" />
+        <register type="NOA" address="0x00009888" value="0x083500C1" />
+        <register type="NOA" address="0x00009888" value="0x0A35C500" />
+        <register type="NOA" address="0x00009888" value="0x0C3500C3" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x1853002A" />
+        <register type="NOA" address="0x00009888" value="0x0A54E000" />
+        <register type="NOA" address="0x00009888" value="0x0C55C500" />
+        <register type="NOA" address="0x00009888" value="0x0E55C1C3" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x00DC8000" />
+        <register type="NOA" address="0x00009888" value="0x02DCC000" />
+        <register type="NOA" address="0x00009888" value="0x04DC4000" />
+        <register type="NOA" address="0x00009888" value="0x04BD8000" />
+        <register type="NOA" address="0x00009888" value="0x06BD8000" />
+        <register type="NOA" address="0x00009888" value="0x02D8C000" />
+        <register type="NOA" address="0x00009888" value="0x02DB8000" />
+        <register type="NOA" address="0x00009888" value="0x04DB4000" />
+        <register type="NOA" address="0x00009888" value="0x06DB4000" />
+        <register type="NOA" address="0x00009888" value="0x08DB8000" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F00AA" />
+        <register type="NOA" address="0x00009888" value="0x02B84000" />
+        <register type="NOA" address="0x00009888" value="0x04B84000" />
+        <register type="NOA" address="0x00009888" value="0x06B84000" />
+        <register type="NOA" address="0x00009888" value="0x08B84000" />
+        <register type="NOA" address="0x00009888" value="0x0AB88000" />
+        <register type="NOA" address="0x00009888" value="0x0CB88000" />
+        <register type="NOA" address="0x00009888" value="0x00B98000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x04B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x0ABA8000" />
+        <register type="NOA" address="0x00009888" value="0x0CBA8000" />
+        <register type="NOA" address="0x00009888" value="0x04938000" />
+        <register type="NOA" address="0x00009888" value="0x06938000" />
+        <register type="NOA" address="0x00009888" value="0x0494C000" />
+        <register type="NOA" address="0x00009888" value="0x0295CFC7" />
+        <register type="NOA" address="0x00009888" value="0x10950000" />
+        <register type="NOA" address="0x00009888" value="0x02B38000" />
+        <register type="NOA" address="0x00009888" value="0x08B38000" />
+        <register type="NOA" address="0x00009888" value="0x04B42000" />
+        <register type="NOA" address="0x00009888" value="0x06B41000" />
+        <register type="NOA" address="0x00009888" value="0x00B5C700" />
+        <register type="NOA" address="0x00009888" value="0x04B500CF" />
+        <register type="NOA" address="0x00009888" value="0x10B50000" />
+        <register type="NOA" address="0x00009888" value="0x0AD38000" />
+        <register type="NOA" address="0x00009888" value="0x0CD38000" />
+        <register type="NOA" address="0x00009888" value="0x06D46000" />
+        <register type="NOA" address="0x00009888" value="0x04D5C700" />
+        <register type="NOA" address="0x00009888" value="0x06D500CF" />
+        <register type="NOA" address="0x00009888" value="0x10D50000" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x258B555A" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x238B5500" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800882" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45801082" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x478014A5" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800002" />
+        <register type="NOA" address="0x00009888" value="0x41800C62" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x0000FE7F" />
+        <register type="OA" address="0x00002780" value="0x00000000" />
+        <register type="OA" address="0x00002784" value="0x0000FF9F" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000FFE7" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFFB" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FFFD" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extra Gen8"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extra"
+       hw_config_guid="8fb61ba2-2fbb-454c-a136-2dec5a8a595e"
+       chipset="BDW"
+       symbol_name="ComputeExtra"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active including Ext Math"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing including Extended Math processing"
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ  C 5 READ C 6 READ FADD C 7 READ FADD C 2 READ FADD C 3 READ FADD C 4 READ FADD 8 FMUL FADD 100 FMUL $EuCoresTotalCount FDIV $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active_adjusted"
+             units="percent"
+             symbol_name="Fpu1ActiveAdjusted"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x161503E0" />
+        <register type="NOA" address="0x00009888" value="0x163503E0" />
+        <register type="NOA" address="0x00009888" value="0x165503E0" />
+        <register type="NOA" address="0x00009888" value="0x169503E0" />
+        <register type="NOA" address="0x00009888" value="0x16B503E0" />
+        <register type="NOA" address="0x00009888" value="0x16D503E0" />
+        <register type="NOA" address="0x00009888" value="0x045CC000" />
+        <register type="NOA" address="0x00009888" value="0x083D8000" />
+        <register type="NOA" address="0x00009888" value="0x04584000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00A8" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0C388000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A8000" />
+        <register type="NOA" address="0x00009888" value="0x08138000" />
+        <register type="NOA" address="0x00009888" value="0x06141000" />
+        <register type="NOA" address="0x00009888" value="0x041500C3" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x0A338000" />
+        <register type="NOA" address="0x00009888" value="0x06342000" />
+        <register type="NOA" address="0x00009888" value="0x0435C300" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x0C538000" />
+        <register type="NOA" address="0x00009888" value="0x06544000" />
+        <register type="NOA" address="0x00009888" value="0x065500C3" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x00DC8000" />
+        <register type="NOA" address="0x00009888" value="0x02DC4000" />
+        <register type="NOA" address="0x00009888" value="0x02BD8000" />
+        <register type="NOA" address="0x00009888" value="0x00D88000" />
+        <register type="NOA" address="0x00009888" value="0x02DB4000" />
+        <register type="NOA" address="0x00009888" value="0x04DB8000" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F0002" />
+        <register type="NOA" address="0x00009888" value="0x02B84000" />
+        <register type="NOA" address="0x00009888" value="0x04B84000" />
+        <register type="NOA" address="0x00009888" value="0x06B88000" />
+        <register type="NOA" address="0x00009888" value="0x00B98000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06BA8000" />
+        <register type="NOA" address="0x00009888" value="0x02938000" />
+        <register type="NOA" address="0x00009888" value="0x04942000" />
+        <register type="NOA" address="0x00009888" value="0x0095C300" />
+        <register type="NOA" address="0x00009888" value="0x10950000" />
+        <register type="NOA" address="0x00009888" value="0x04B38000" />
+        <register type="NOA" address="0x00009888" value="0x04B44000" />
+        <register type="NOA" address="0x00009888" value="0x02B500C3" />
+        <register type="NOA" address="0x00009888" value="0x10B50000" />
+        <register type="NOA" address="0x00009888" value="0x06D38000" />
+        <register type="NOA" address="0x00009888" value="0x04D48000" />
+        <register type="NOA" address="0x00009888" value="0x02D5C300" />
+        <register type="NOA" address="0x00009888" value="0x10D50000" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B3500" />
+        <register type="NOA" address="0x00009888" value="0x258B0005" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x2185000A" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800C40" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41801482" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00001000" />
+        <register type="FLEX" address="0x0000E558" value="0x00003002" />
+        <register type="FLEX" address="0x0000E658" value="0x00005004" />
+        <register type="FLEX" address="0x0000E758" value="0x00011010" />
+        <register type="FLEX" address="0x0000E45C" value="0x00050012" />
+        <register type="FLEX" address="0x0000E55C" value="0x00052051" />
+        <register type="FLEX" address="0x0000E65C" value="0x00000008" />
+    </register_config>
+  </set>
+
+  <set name="Media Vme Pipe Gen8"
+       mdapi_supported_apis="MEDIA IO BB"
+       underscore_name="vme_pipe"
+       hw_config_guid="e1743ca0-7fc8-410b-a066-de7bbb9280b7"
+       chipset="BDW"
+       symbol_name="VMEPipe"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL  GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="8 A 10 READ FMUL $EuThreadsCount FDIV $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VME Busy"
+             description="The percentage of time in which VME (IME or CRE) was actively processing data."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ B 3 READ FADD 2 FDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vme_busy"
+             units="percent"
+             symbol_name="VMEBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Overview System Batch Tier2"
+             mdapi_group="VME Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x14100812" />
+        <register type="NOA" address="0x00009888" value="0x14125800" />
+        <register type="NOA" address="0x00009888" value="0x161200C0" />
+        <register type="NOA" address="0x00009888" value="0x14300812" />
+        <register type="NOA" address="0x00009888" value="0x14325800" />
+        <register type="NOA" address="0x00009888" value="0x163200C0" />
+        <register type="NOA" address="0x00009888" value="0x005C4000" />
+        <register type="NOA" address="0x00009888" value="0x065C8000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5CC000" />
+        <register type="NOA" address="0x00009888" value="0x003D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E3D8000" />
+        <register type="NOA" address="0x00009888" value="0x183D2800" />
+        <register type="NOA" address="0x00009888" value="0x00584000" />
+        <register type="NOA" address="0x00009888" value="0x06588000" />
+        <register type="NOA" address="0x00009888" value="0x0858C000" />
+        <register type="NOA" address="0x00009888" value="0x005B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x185B9400" />
+        <register type="NOA" address="0x00009888" value="0x1A5B002A" />
+        <register type="NOA" address="0x00009888" value="0x0C1F0800" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA00" />
+        <register type="NOA" address="0x00009888" value="0x101F002A" />
+        <register type="NOA" address="0x00009888" value="0x00384000" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18380155" />
+        <register type="NOA" address="0x00009888" value="0x00392000" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x00100047" />
+        <register type="NOA" address="0x00009888" value="0x06101A80" />
+        <register type="NOA" address="0x00009888" value="0x10100000" />
+        <register type="NOA" address="0x00009888" value="0x0810C000" />
+        <register type="NOA" address="0x00009888" value="0x0811C000" />
+        <register type="NOA" address="0x00009888" value="0x08126151" />
+        <register type="NOA" address="0x00009888" value="0x10120000" />
+        <register type="NOA" address="0x00009888" value="0x00134000" />
+        <register type="NOA" address="0x00009888" value="0x0E134000" />
+        <register type="NOA" address="0x00009888" value="0x161300A0" />
+        <register type="NOA" address="0x00009888" value="0x0A301AC7" />
+        <register type="NOA" address="0x00009888" value="0x10300000" />
+        <register type="NOA" address="0x00009888" value="0x0C30C000" />
+        <register type="NOA" address="0x00009888" value="0x0C31C000" />
+        <register type="NOA" address="0x00009888" value="0x0C326151" />
+        <register type="NOA" address="0x00009888" value="0x10320000" />
+        <register type="NOA" address="0x00009888" value="0x16332A00" />
+        <register type="NOA" address="0x00009888" value="0x18330001" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A2AA0" />
+        <register type="NOA" address="0x00009888" value="0x238B0020" />
+        <register type="NOA" address="0x00009888" value="0x258B5550" />
+        <register type="NOA" address="0x00009888" value="0x278B0001" />
+        <register type="NOA" address="0x00009888" value="0x1F850080" />
+        <register type="NOA" address="0x00009888" value="0x2185AAA0" />
+        <register type="NOA" address="0x00009888" value="0x23850002" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830015" />
+        <register type="NOA" address="0x00009888" value="0x01844000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x11804000" />
+        <register type="NOA" address="0x00009888" value="0x17808000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3D800800" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800002" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800884" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800002" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00100030" />
+        <register type="OA" address="0x00002774" value="0x0000FFF9" />
+        <register type="OA" address="0x00002778" value="0x00000002" />
+        <register type="OA" address="0x0000277C" value="0x0000FFFC" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFF3" />
+        <register type="OA" address="0x00002788" value="0x00100180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FF3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00008003" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen8"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="d6de6f55-e526-4f79-a6a6-d7315c09044e"
+       chipset="BDW"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x198B0000" />
+        <register type="NOA" address="0x00009888" value="0x078B0066" />
+        <register type="NOA" address="0x00009888" value="0x118B0000" />
+        <register type="NOA" address="0x00009888" value="0x258B0000" />
+        <register type="NOA" address="0x00009888" value="0x21850008" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
diff --git a/src/mesa/drivers/dri/i965/brw_oa_bxt.xml b/src/mesa/drivers/dri/i965/brw_oa_bxt.xml
new file mode 100644
index 0000000..18d14d9
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_bxt.xml
@@ -0,0 +1,9211 @@
+<?xml version="1.0"?>
+<metrics version="1491575452" merge_md5="">
+  <set name="Render Metrics Basic Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="22b9519a-e9ba-4c41-8b54-f4f8ca14fa0a"
+       chipset="BXT"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Cache Misses"
+             description="The total number of sampler cache misses in all LODs in all sampler units."
+             data_type="uint64"
+             equation="B 4 READ 8 UMUL"
+             underscore_name="sampler_l1_misses"
+             units="messages"
+             symbol_name="SamplerL1Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 1 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$SamplerL1Misses 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="$Sampler0Busy $Sampler1Busy FMAX"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="L3 Lookup Accesses w/o IC"
+             description="The total number of L3 cache lookup accesses w/o IC."
+             data_type="uint64"
+             equation="$SamplerL1Misses $ShaderMemoryAccesses UADD"
+             underscore_name="l3_lookups"
+             units="messages"
+             symbol_name="L3Lookups"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x03 UGTE"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x166C00F0" />
+        <register type="NOA" address="0x00009888" value="0x12120280" />
+        <register type="NOA" address="0x00009888" value="0x12320280" />
+        <register type="NOA" address="0x00009888" value="0x11930317" />
+        <register type="NOA" address="0x00009888" value="0x159303DF" />
+        <register type="NOA" address="0x00009888" value="0x3F900C00" />
+        <register type="NOA" address="0x00009888" value="0x419000A0" />
+        <register type="NOA" address="0x00009888" value="0x002D1000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x082D5000" />
+        <register type="NOA" address="0x00009888" value="0x0A2D1000" />
+        <register type="NOA" address="0x00009888" value="0x0C2E0800" />
+        <register type="NOA" address="0x00009888" value="0x0E2E5900" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4C4000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E2000" />
+        <register type="NOA" address="0x00009888" value="0x1C4F0010" />
+        <register type="NOA" address="0x00009888" value="0x0A6C0053" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A0FCC00" />
+        <register type="NOA" address="0x00009888" value="0x1C0F0002" />
+        <register type="NOA" address="0x00009888" value="0x1C2C0040" />
+        <register type="NOA" address="0x00009888" value="0x00101000" />
+        <register type="NOA" address="0x00009888" value="0x04101000" />
+        <register type="NOA" address="0x00009888" value="0x00114000" />
+        <register type="NOA" address="0x00009888" value="0x08114000" />
+        <register type="NOA" address="0x00009888" value="0x00120020" />
+        <register type="NOA" address="0x00009888" value="0x08120021" />
+        <register type="NOA" address="0x00009888" value="0x00141000" />
+        <register type="NOA" address="0x00009888" value="0x08141000" />
+        <register type="NOA" address="0x00009888" value="0x02308000" />
+        <register type="NOA" address="0x00009888" value="0x04302000" />
+        <register type="NOA" address="0x00009888" value="0x06318000" />
+        <register type="NOA" address="0x00009888" value="0x08318000" />
+        <register type="NOA" address="0x00009888" value="0x06320800" />
+        <register type="NOA" address="0x00009888" value="0x08320840" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x06344000" />
+        <register type="NOA" address="0x00009888" value="0x08344000" />
+        <register type="NOA" address="0x00009888" value="0x0D931831" />
+        <register type="NOA" address="0x00009888" value="0x0F939F3F" />
+        <register type="NOA" address="0x00009888" value="0x01939E80" />
+        <register type="NOA" address="0x00009888" value="0x039303BC" />
+        <register type="NOA" address="0x00009888" value="0x0593000E" />
+        <register type="NOA" address="0x00009888" value="0x1993002A" />
+        <register type="NOA" address="0x00009888" value="0x07930000" />
+        <register type="NOA" address="0x00009888" value="0x09930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900177" />
+        <register type="NOA" address="0x00009888" value="0x1F900187" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25904000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x53901110" />
+        <register type="NOA" address="0x00009888" value="0x43900423" />
+        <register type="NOA" address="0x00009888" value="0x55900111" />
+        <register type="NOA" address="0x00009888" value="0x47900C02" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900020" />
+        <register type="NOA" address="0x00009888" value="0x59901111" />
+        <register type="NOA" address="0x00009888" value="0x4B900421" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900001" />
+        <register type="NOA" address="0x00009888" value="0x45900821" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="012d72cf-82a9-4d25-8ddf-74076fd30797"
+       chipset="BXT"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 3 READ B 4 READ UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 1 READ C 2 READ UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL  $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x124F1C00" />
+        <register type="NOA" address="0x00009888" value="0x39900340" />
+        <register type="NOA" address="0x00009888" value="0x3F900C00" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x002D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x082D4000" />
+        <register type="NOA" address="0x00009888" value="0x0A2D1000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C2E1400" />
+        <register type="NOA" address="0x00009888" value="0x0E2E5100" />
+        <register type="NOA" address="0x00009888" value="0x102E0114" />
+        <register type="NOA" address="0x00009888" value="0x044CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4C4000" />
+        <register type="NOA" address="0x00009888" value="0x104C8000" />
+        <register type="NOA" address="0x00009888" value="0x124C8000" />
+        <register type="NOA" address="0x00009888" value="0x164C2000" />
+        <register type="NOA" address="0x00009888" value="0x004EA000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4EA000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x004F6B42" />
+        <register type="NOA" address="0x00009888" value="0x064F6200" />
+        <register type="NOA" address="0x00009888" value="0x084F4100" />
+        <register type="NOA" address="0x00009888" value="0x0A4F0061" />
+        <register type="NOA" address="0x00009888" value="0x0C4F6C4C" />
+        <register type="NOA" address="0x00009888" value="0x0E4F4B00" />
+        <register type="NOA" address="0x00009888" value="0x1A4F0000" />
+        <register type="NOA" address="0x00009888" value="0x1C4F0000" />
+        <register type="NOA" address="0x00009888" value="0x180F5000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F8800" />
+        <register type="NOA" address="0x00009888" value="0x1C0F08A2" />
+        <register type="NOA" address="0x00009888" value="0x182C4000" />
+        <register type="NOA" address="0x00009888" value="0x1C2C1451" />
+        <register type="NOA" address="0x00009888" value="0x1E2C0001" />
+        <register type="NOA" address="0x00009888" value="0x1A2C0010" />
+        <register type="NOA" address="0x00009888" value="0x01938000" />
+        <register type="NOA" address="0x00009888" value="0x0F938000" />
+        <register type="NOA" address="0x00009888" value="0x19938A28" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x19900177" />
+        <register type="NOA" address="0x00009888" value="0x1B900178" />
+        <register type="NOA" address="0x00009888" value="0x1D900125" />
+        <register type="NOA" address="0x00009888" value="0x1F900123" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x25904000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x53901000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x55900111" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="ce416533-e49e-4211-80af-ec513590a914"
+       chipset="BXT"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x0C2E001F" />
+        <register type="NOA" address="0x00009888" value="0x0A2F0000" />
+        <register type="NOA" address="0x00009888" value="0x10186800" />
+        <register type="NOA" address="0x00009888" value="0x11810019" />
+        <register type="NOA" address="0x00009888" value="0x15810013" />
+        <register type="NOA" address="0x00009888" value="0x13820020" />
+        <register type="NOA" address="0x00009888" value="0x11830020" />
+        <register type="NOA" address="0x00009888" value="0x17840000" />
+        <register type="NOA" address="0x00009888" value="0x11860007" />
+        <register type="NOA" address="0x00009888" value="0x21860000" />
+        <register type="NOA" address="0x00009888" value="0x178703E0" />
+        <register type="NOA" address="0x00009888" value="0x0C2D8000" />
+        <register type="NOA" address="0x00009888" value="0x042D4000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x022E5400" />
+        <register type="NOA" address="0x00009888" value="0x002E0000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0080" />
+        <register type="NOA" address="0x00009888" value="0x082F0040" />
+        <register type="NOA" address="0x00009888" value="0x002F0000" />
+        <register type="NOA" address="0x00009888" value="0x06143000" />
+        <register type="NOA" address="0x00009888" value="0x06174000" />
+        <register type="NOA" address="0x00009888" value="0x06180012" />
+        <register type="NOA" address="0x00009888" value="0x00180000" />
+        <register type="NOA" address="0x00009888" value="0x0D804000" />
+        <register type="NOA" address="0x00009888" value="0x0F804000" />
+        <register type="NOA" address="0x00009888" value="0x05804000" />
+        <register type="NOA" address="0x00009888" value="0x09810200" />
+        <register type="NOA" address="0x00009888" value="0x0B810030" />
+        <register type="NOA" address="0x00009888" value="0x03810003" />
+        <register type="NOA" address="0x00009888" value="0x21819140" />
+        <register type="NOA" address="0x00009888" value="0x23819050" />
+        <register type="NOA" address="0x00009888" value="0x25810018" />
+        <register type="NOA" address="0x00009888" value="0x0B820980" />
+        <register type="NOA" address="0x00009888" value="0x03820D80" />
+        <register type="NOA" address="0x00009888" value="0x11820000" />
+        <register type="NOA" address="0x00009888" value="0x0182C000" />
+        <register type="NOA" address="0x00009888" value="0x07828000" />
+        <register type="NOA" address="0x00009888" value="0x09824000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x0D830004" />
+        <register type="NOA" address="0x00009888" value="0x0583000C" />
+        <register type="NOA" address="0x00009888" value="0x0F831000" />
+        <register type="NOA" address="0x00009888" value="0x01848072" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x07848000" />
+        <register type="NOA" address="0x00009888" value="0x09844000" />
+        <register type="NOA" address="0x00009888" value="0x0F848000" />
+        <register type="NOA" address="0x00009888" value="0x07860000" />
+        <register type="NOA" address="0x00009888" value="0x09860092" />
+        <register type="NOA" address="0x00009888" value="0x0F860400" />
+        <register type="NOA" address="0x00009888" value="0x01869100" />
+        <register type="NOA" address="0x00009888" value="0x0F870065" />
+        <register type="NOA" address="0x00009888" value="0x01870000" />
+        <register type="NOA" address="0x00009888" value="0x19930800" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x1B952000" />
+        <register type="NOA" address="0x00009888" value="0x1D955055" />
+        <register type="NOA" address="0x00009888" value="0x1F951455" />
+        <register type="NOA" address="0x00009888" value="0x0992A000" />
+        <register type="NOA" address="0x00009888" value="0x0F928000" />
+        <register type="NOA" address="0x00009888" value="0x1192A800" />
+        <register type="NOA" address="0x00009888" value="0x1392028A" />
+        <register type="NOA" address="0x00009888" value="0x0B92A000" />
+        <register type="NOA" address="0x00009888" value="0x0D922000" />
+        <register type="NOA" address="0x00009888" value="0x13908000" />
+        <register type="NOA" address="0x00009888" value="0x21908000" />
+        <register type="NOA" address="0x00009888" value="0x23908000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27908000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x2F908000" />
+        <register type="NOA" address="0x00009888" value="0x31908000" />
+        <register type="NOA" address="0x00009888" value="0x15908000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900C01" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900863" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900061" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C22" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="398e2452-18d7-42d0-b241-e4d0a9148ada"
+       chipset="BXT"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank0Reads"
+             description="The total number of GTI memory reads from L3 Bank 0 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_reads"
+             units="messages"
+             symbol_name="GtiL3Bank0Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all accesses from GTI to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiL3Bank3Reads"
+             description="The total number of GTI memory reads from L3 Bank 3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_reads"
+             units="messages"
+             symbol_name="GtiL3Bank3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiRsMemoryReads"
+             description="The total number of GTI memory reads from Resource Streamer."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_rs_memory_reads"
+             units="messages"
+             symbol_name="GtiRsMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiHizMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache (Hi-Depth Cache misses)."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_reads"
+             units="messages"
+             symbol_name="GtiHizMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="GtiRccMemoryReads"
+             description="The total number of GTI memory reads from Render Color Cache (Render Color Cache misses)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_reads"
+             units="messages"
+             symbol_name="GtiRccMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank1Reads"
+             description="The total number of GTI memory reads from L3 Bank 1 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_reads"
+             units="messages"
+             symbol_name="GtiL3Bank1Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiCmdStreamerMemoryReads"
+             description="The total number of GTI memory reads from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_reads"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GtiL3Bank2Reads"
+             description="The total number of GTI memory reads from L3 Bank 2 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_reads"
+             units="messages"
+             symbol_name="GtiL3Bank2Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiMemoryReads"
+             description="The total number of GTI memory reads."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache (Render Depth Cache misses)."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="GtiMscMemoryReads"
+             description="The total number of GTI memory reads from Multisampling Color Cache (Multisampling Color Cache misses)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_reads"
+             units="messages"
+             symbol_name="GtiMscMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiVfMemoryReads"
+             description="The total number of GTI memory reads from Vertex Fetch."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="gti_vf_memory_reads"
+             units="messages"
+             symbol_name="GtiVfMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache (Stencil Cache misses)."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Reads $GtiL3Bank1Reads $GtiL3Bank2Reads $GtiL3Bank3Reads UADD UADD UADD"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x19800343" />
+        <register type="NOA" address="0x00009888" value="0x39900340" />
+        <register type="NOA" address="0x00009888" value="0x3F901000" />
+        <register type="NOA" address="0x00009888" value="0x41900003" />
+        <register type="NOA" address="0x00009888" value="0x03803180" />
+        <register type="NOA" address="0x00009888" value="0x058035E2" />
+        <register type="NOA" address="0x00009888" value="0x0780006A" />
+        <register type="NOA" address="0x00009888" value="0x11800000" />
+        <register type="NOA" address="0x00009888" value="0x2181A000" />
+        <register type="NOA" address="0x00009888" value="0x2381000A" />
+        <register type="NOA" address="0x00009888" value="0x1D950550" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D92A000" />
+        <register type="NOA" address="0x00009888" value="0x0F922000" />
+        <register type="NOA" address="0x00009888" value="0x13900170" />
+        <register type="NOA" address="0x00009888" value="0x21900171" />
+        <register type="NOA" address="0x00009888" value="0x23900172" />
+        <register type="NOA" address="0x00009888" value="0x25900173" />
+        <register type="NOA" address="0x00009888" value="0x27900174" />
+        <register type="NOA" address="0x00009888" value="0x29900175" />
+        <register type="NOA" address="0x00009888" value="0x2B900176" />
+        <register type="NOA" address="0x00009888" value="0x2D900177" />
+        <register type="NOA" address="0x00009888" value="0x2F90017F" />
+        <register type="NOA" address="0x00009888" value="0x31900125" />
+        <register type="NOA" address="0x00009888" value="0x15900123" />
+        <register type="NOA" address="0x00009888" value="0x17900121" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43901084" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47901080" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49901084" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B901084" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900004" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F872" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="d324a0d6-7269-4847-a5c2-6f71ddc7fed5"
+       chipset="BXT"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiMemoryWrites"
+             description="The total number of GTI memory writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all GTI accesses to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiMscMemoryWrites"
+             description="The total number of GTI memory writes from Multisampling Color Cache (Multisampling Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_writes"
+             units="messages"
+             symbol_name="GtiMscMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiCmdStreamerMemoryWrites"
+             description="The total number of GTI memory writes from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_writes"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiL3Bank0Writes"
+             description="The total number of GTI memory writes from L3 Bank 0 (L3 Bank 0 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_writes"
+             units="messages"
+             symbol_name="GtiL3Bank0Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank1Writes"
+             description="The total number of GTI memory writes from L3 Bank 1 (L3 Bank 1 invalidations)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_writes"
+             units="messages"
+             symbol_name="GtiL3Bank1Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank2Writes"
+             description="The total number of GTI memory writes from L3 Bank 2 (L3 Bank 2 invalidations)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_writes"
+             units="messages"
+             symbol_name="GtiL3Bank2Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank3Writes"
+             description="The total number of GTI memory writes from L3 Bank 3 (L3 Bank 3 invalidations)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_writes"
+             units="messages"
+             symbol_name="GtiL3Bank3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Writes $GtiL3Bank1Writes $GtiL3Bank2Writes $GtiL3Bank3Writes UADD UADD UADD"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRccMemoryWrites"
+             description="The total number of GTI memory writes from Render Color Cache (Render Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_writes"
+             units="messages"
+             symbol_name="GtiRccMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiSoMemoryWrites"
+             description="The total number of GTI memory writes from Stream Output."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_so_memory_writes"
+             units="messages"
+             symbol_name="GtiSoMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiStcMemoryWrites"
+             description="The total number of GTI memory writes from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_writes"
+             units="messages"
+             symbol_name="GtiStcMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GtiRczMemoryWrites"
+             description="The total number of GTI memory writes from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_writes"
+             units="messages"
+             symbol_name="GtiRczMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiHizMemoryWrites"
+             description="The total number of GTI memory writes from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_writes"
+             units="messages"
+             symbol_name="GtiHizMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x19800343" />
+        <register type="NOA" address="0x00009888" value="0x39900340" />
+        <register type="NOA" address="0x00009888" value="0x3F900000" />
+        <register type="NOA" address="0x00009888" value="0x41900080" />
+        <register type="NOA" address="0x00009888" value="0x03803180" />
+        <register type="NOA" address="0x00009888" value="0x058035E2" />
+        <register type="NOA" address="0x00009888" value="0x0780006A" />
+        <register type="NOA" address="0x00009888" value="0x11800000" />
+        <register type="NOA" address="0x00009888" value="0x2181A000" />
+        <register type="NOA" address="0x00009888" value="0x2381000A" />
+        <register type="NOA" address="0x00009888" value="0x1D950550" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D92A000" />
+        <register type="NOA" address="0x00009888" value="0x0F922000" />
+        <register type="NOA" address="0x00009888" value="0x13900180" />
+        <register type="NOA" address="0x00009888" value="0x21900181" />
+        <register type="NOA" address="0x00009888" value="0x23900182" />
+        <register type="NOA" address="0x00009888" value="0x25900183" />
+        <register type="NOA" address="0x00009888" value="0x27900184" />
+        <register type="NOA" address="0x00009888" value="0x29900185" />
+        <register type="NOA" address="0x00009888" value="0x2B900186" />
+        <register type="NOA" address="0x00009888" value="0x2D900187" />
+        <register type="NOA" address="0x00009888" value="0x2F900170" />
+        <register type="NOA" address="0x00009888" value="0x31900125" />
+        <register type="NOA" address="0x00009888" value="0x15900123" />
+        <register type="NOA" address="0x00009888" value="0x17900121" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43901084" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47901080" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49901084" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B901084" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900004" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F822" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="caf3596a-7bb1-4dec-b3b3-2a080d283b49"
+       chipset="BXT"
+       symbol_name="ComputeExtended"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Writes 0"
+             description="The subslice 0 typed writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="typed_writes0"
+             units="messages"
+             symbol_name="TypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuTypedAtomics0"
+             description="The subslice 0 EU Typed Atomics subslice 0."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_typed_atomics0"
+             units="messages"
+             symbol_name="EuTypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Atomics 0"
+             description="The subslice 0 typed atomics."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="typed_atomics0"
+             units="messages"
+             symbol_name="TypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedAtomicsPerCacheLine"
+             description="The ratio of EU typed atomics requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedAtomics0 $TypedAtomics0 FDIV"
+             underscore_name="typed_atomics_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedAtomicsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedReads0"
+             description="The subslice 0 EU Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_untyped_reads0"
+             units="messages"
+             symbol_name="EuUntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Writes 0"
+             description="The subslice 0 untyped writes (including SLM writes)."
+             data_type="uint64"
+             equation="C 1 READ"
+             underscore_name="untyped_writes0"
+             units="messages"
+             symbol_name="UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedAtomics0"
+             description="The subslice 0 EU Untyped Atomics subslice 0."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_untyped_atomics0"
+             units="messages"
+             symbol_name="EuUntypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedWrites0"
+             description="The subslice 0 EU Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="eu_untyped_writes0"
+             units="messages"
+             symbol_name="EuUntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedWrites0"
+             description="The subslice 0 EU A64 Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_a64_untyped_writes0"
+             units="messages"
+             symbol_name="EuA64UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedWritesPerCacheLine"
+             description="The ratio of EU untyped write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuUntypedWrites0 $EuA64UntypedWrites0 UADD $UntypedWrites0 FDIV"
+             underscore_name="untyped_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedWrites0"
+             description="The subslice 0 EU Typed Writes subslice 0."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="eu_typed_writes0"
+             units="messages"
+             symbol_name="EuTypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedWritesPerCacheLine"
+             description="The ratio of EU typed write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedWrites0 $TypedWrites0 FDIV"
+             underscore_name="typed_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Reads 0"
+             description="The subslice 0 typed reads."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="typed_reads0"
+             units="messages"
+             symbol_name="TypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Reads 0"
+             description="The subslice 0 untyped reads (including SLM reads)."
+             data_type="uint64"
+             equation="C 3 READ"
+             underscore_name="untyped_reads0"
+             units="messages"
+             symbol_name="UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedReads0"
+             description="The subslice 0 EU A64 Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_a64_untyped_reads0"
+             units="messages"
+             symbol_name="EuA64UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedReads0"
+             description="The subslice 0 EU Typed Reads subslice 0."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="eu_typed_reads0"
+             units="messages"
+             symbol_name="EuTypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedReadsPerCacheLine"
+             description="The ratio of EU untyped read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuUntypedReads0 $EuA64UntypedReads0 UADD $UntypedReads0 FDIV"
+             underscore_name="untyped_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="The ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuTypedReads0 $TypedReads0 FDIV"
+             underscore_name="typed_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x141C0160" />
+        <register type="NOA" address="0x00009888" value="0x161C0015" />
+        <register type="NOA" address="0x00009888" value="0x181C0120" />
+        <register type="NOA" address="0x00009888" value="0x002D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x082D5000" />
+        <register type="NOA" address="0x00009888" value="0x0A2D5000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D5000" />
+        <register type="NOA" address="0x00009888" value="0x022D5000" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x0C2E5400" />
+        <register type="NOA" address="0x00009888" value="0x0E2E5515" />
+        <register type="NOA" address="0x00009888" value="0x102E0155" />
+        <register type="NOA" address="0x00009888" value="0x044CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4CC000" />
+        <register type="NOA" address="0x00009888" value="0x0E4CC000" />
+        <register type="NOA" address="0x00009888" value="0x104C8000" />
+        <register type="NOA" address="0x00009888" value="0x124C8000" />
+        <register type="NOA" address="0x00009888" value="0x144C8000" />
+        <register type="NOA" address="0x00009888" value="0x164C2000" />
+        <register type="NOA" address="0x00009888" value="0x064CC000" />
+        <register type="NOA" address="0x00009888" value="0x084CC000" />
+        <register type="NOA" address="0x00009888" value="0x004EA000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084EA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4EA000" />
+        <register type="NOA" address="0x00009888" value="0x0C4EA000" />
+        <register type="NOA" address="0x00009888" value="0x0E4EA000" />
+        <register type="NOA" address="0x00009888" value="0x024EA000" />
+        <register type="NOA" address="0x00009888" value="0x044EA000" />
+        <register type="NOA" address="0x00009888" value="0x0E4F4B41" />
+        <register type="NOA" address="0x00009888" value="0x004F4200" />
+        <register type="NOA" address="0x00009888" value="0x024F404C" />
+        <register type="NOA" address="0x00009888" value="0x1C4F0000" />
+        <register type="NOA" address="0x00009888" value="0x1A4F0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x061B8000" />
+        <register type="NOA" address="0x00009888" value="0x081BC000" />
+        <register type="NOA" address="0x00009888" value="0x0A1BC000" />
+        <register type="NOA" address="0x00009888" value="0x0C1BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x001C0031" />
+        <register type="NOA" address="0x00009888" value="0x061C1900" />
+        <register type="NOA" address="0x00009888" value="0x081C1A33" />
+        <register type="NOA" address="0x00009888" value="0x0A1C1B35" />
+        <register type="NOA" address="0x00009888" value="0x0C1C3337" />
+        <register type="NOA" address="0x00009888" value="0x041C31C7" />
+        <register type="NOA" address="0x00009888" value="0x180F5000" />
+        <register type="NOA" address="0x00009888" value="0x1A0FA8AA" />
+        <register type="NOA" address="0x00009888" value="0x1C0F0AAA" />
+        <register type="NOA" address="0x00009888" value="0x182C8000" />
+        <register type="NOA" address="0x00009888" value="0x1C2C6AAA" />
+        <register type="NOA" address="0x00009888" value="0x1E2C0001" />
+        <register type="NOA" address="0x00009888" value="0x1A2C2950" />
+        <register type="NOA" address="0x00009888" value="0x01938000" />
+        <register type="NOA" address="0x00009888" value="0x0F938000" />
+        <register type="NOA" address="0x00009888" value="0x1993AAAA" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x05938000" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25904000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x29904000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900420" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900400" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900001" />
+        <register type="NOA" address="0x00009888" value="0x45900001" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FC2A" />
+        <register type="OA" address="0x00002774" value="0x0000BF00" />
+        <register type="OA" address="0x00002778" value="0x0007FC6A" />
+        <register type="OA" address="0x0000277C" value="0x0000BF00" />
+        <register type="OA" address="0x00002780" value="0x0007FC92" />
+        <register type="OA" address="0x00002784" value="0x0000BF00" />
+        <register type="OA" address="0x00002788" value="0x0007FCA2" />
+        <register type="OA" address="0x0000278C" value="0x0000BF00" />
+        <register type="OA" address="0x00002790" value="0x0007FC32" />
+        <register type="OA" address="0x00002794" value="0x0000BF00" />
+        <register type="OA" address="0x00002798" value="0x0007FC9A" />
+        <register type="OA" address="0x0000279C" value="0x0000BF00" />
+        <register type="OA" address="0x000027A0" value="0x0007FE6A" />
+        <register type="OA" address="0x000027A4" value="0x0000BF00" />
+        <register type="OA" address="0x000027A8" value="0x0007FE7A" />
+        <register type="OA" address="0x000027AC" value="0x0000BF00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics L3 Cache Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_l3_cache"
+       hw_config_guid="49b956e2-d5b9-47e0-9d8a-cee5e8cec527"
+       chipset="BXT"
+       symbol_name="ComputeL3Cache"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 03 Accesses"
+             description="The total number of accesses to L3 Bank 03."
+             data_type="uint64"
+             equation="B 3 READ 2 UMUL"
+             underscore_name="l3_bank03_accesses"
+             units="messages"
+             symbol_name="L3Bank03Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Accesses"
+             description="The total number of L3 accesses from all entities."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ B 2 READ B 3 READ UADD UADD UADD 2 UMUL"
+             underscore_name="l3_accesses"
+             units="messages"
+             symbol_name="L3Accesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 29 READ 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="EU FPU0 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu0_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ C 5 READ UADD"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="L3 Bank 00 Accesses"
+             description="The total number of accesses to L3 Bank 00."
+             data_type="uint64"
+             equation="C 0 READ 2 UMUL"
+             underscore_name="l3_bank00_accesses"
+             units="messages"
+             symbol_name="L3Bank00Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU0 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 19 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu0_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 14 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu1_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 00 IC Accesses"
+             description="The total number of accesses to L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 0 READ B 1 READ UADD 2 UMUL $L3Bank00Accesses UMIN"
+             underscore_name="l3_bank00_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank00IcAccesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 00 IC Hits"
+             description="The total number of hits in L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 1 READ 2 UMUL $L3Bank00IcAccesses UMIN"
+             underscore_name="l3_bank00_ic_hits"
+             units="messages"
+             symbol_name="L3Bank00IcHits"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="L3 Bank 01 Accesses"
+             description="The total number of accesses to L3 Bank 01."
+             data_type="uint64"
+             equation="C 1 READ 2 UMUL"
+             underscore_name="l3_bank01_accesses"
+             units="messages"
+             symbol_name="L3Bank01Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 20 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu1_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="L3 Bank 02 Accesses"
+             description="The total number of accesses to L3 Bank 02."
+             data_type="uint64"
+             equation="B 2 READ 2 UMUL"
+             underscore_name="l3_bank02_accesses"
+             units="messages"
+             symbol_name="L3Bank02Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="L3 Total Throughput"
+             description="The total number of GPU memory bytes transferred via L3."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Accesses 64 UMUL"
+             underscore_name="l3_total_throughput"
+             units="bytes"
+             symbol_name="L3TotalThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="EU FPU1 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C03B0" />
+        <register type="NOA" address="0x00009888" value="0x1593001E" />
+        <register type="NOA" address="0x00009888" value="0x3F900C00" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x002D1000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x082D5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D5000" />
+        <register type="NOA" address="0x00009888" value="0x0C2E0400" />
+        <register type="NOA" address="0x00009888" value="0x0E2E1500" />
+        <register type="NOA" address="0x00009888" value="0x102E0140" />
+        <register type="NOA" address="0x00009888" value="0x044C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4CC000" />
+        <register type="NOA" address="0x00009888" value="0x144C8000" />
+        <register type="NOA" address="0x00009888" value="0x164C2000" />
+        <register type="NOA" address="0x00009888" value="0x004E2000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084EA000" />
+        <register type="NOA" address="0x00009888" value="0x0E4EA000" />
+        <register type="NOA" address="0x00009888" value="0x1A4F4001" />
+        <register type="NOA" address="0x00009888" value="0x1C4F5005" />
+        <register type="NOA" address="0x00009888" value="0x006C0051" />
+        <register type="NOA" address="0x00009888" value="0x066C5000" />
+        <register type="NOA" address="0x00009888" value="0x086C5C5D" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5E5F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x146C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x180F1000" />
+        <register type="NOA" address="0x00009888" value="0x1A0FA800" />
+        <register type="NOA" address="0x00009888" value="0x1C0F0A00" />
+        <register type="NOA" address="0x00009888" value="0x182C4000" />
+        <register type="NOA" address="0x00009888" value="0x1C2C4015" />
+        <register type="NOA" address="0x00009888" value="0x1E2C0001" />
+        <register type="NOA" address="0x00009888" value="0x03931980" />
+        <register type="NOA" address="0x00009888" value="0x05930032" />
+        <register type="NOA" address="0x00009888" value="0x11930000" />
+        <register type="NOA" address="0x00009888" value="0x01938000" />
+        <register type="NOA" address="0x00009888" value="0x0F938000" />
+        <register type="NOA" address="0x00009888" value="0x1993A00A" />
+        <register type="NOA" address="0x00009888" value="0x07930000" />
+        <register type="NOA" address="0x00009888" value="0x09930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900177" />
+        <register type="NOA" address="0x00009888" value="0x1F900178" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x53901000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x55900111" />
+        <register type="NOA" address="0x00009888" value="0x47900001" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900000" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFFA" />
+        <register type="OA" address="0x00002774" value="0x0000FEFE" />
+        <register type="OA" address="0x00002778" value="0x0007FFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000FEFD" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000FBEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000FBDF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00101100" />
+        <register type="FLEX" address="0x0000E45C" value="0x00201200" />
+        <register type="FLEX" address="0x0000E55C" value="0x00301300" />
+        <register type="FLEX" address="0x0000E65C" value="0x00401400" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="f64ef50a-bdba-4b35-8f09-203c13d8ee5a"
+       chipset="BXT"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F0232" />
+        <register type="NOA" address="0x00009888" value="0x124F4640" />
+        <register type="NOA" address="0x00009888" value="0x11834400" />
+        <register type="NOA" address="0x00009888" value="0x022D4000" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0055" />
+        <register type="NOA" address="0x00009888" value="0x064C8000" />
+        <register type="NOA" address="0x00009888" value="0x084CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C4000" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044EA000" />
+        <register type="NOA" address="0x00009888" value="0x064E2000" />
+        <register type="NOA" address="0x00009888" value="0x024F6100" />
+        <register type="NOA" address="0x00009888" value="0x044F416B" />
+        <register type="NOA" address="0x00009888" value="0x064F004B" />
+        <register type="NOA" address="0x00009888" value="0x1A4F0000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F02A8" />
+        <register type="NOA" address="0x00009888" value="0x1A2C5500" />
+        <register type="NOA" address="0x00009888" value="0x0F808000" />
+        <register type="NOA" address="0x00009888" value="0x25810020" />
+        <register type="NOA" address="0x00009888" value="0x0F8305C0" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x1F951000" />
+        <register type="NOA" address="0x00009888" value="0x13920200" />
+        <register type="NOA" address="0x00009888" value="0x31908000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="00ad5a41-7eab-4f7a-9103-49d411c67219"
+       chipset="BXT"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x03 UGTE"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x12643400" />
+        <register type="NOA" address="0x00009888" value="0x12653400" />
+        <register type="NOA" address="0x00009888" value="0x106C6800" />
+        <register type="NOA" address="0x00009888" value="0x126C001E" />
+        <register type="NOA" address="0x00009888" value="0x166C0010" />
+        <register type="NOA" address="0x00009888" value="0x0C2D5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D5000" />
+        <register type="NOA" address="0x00009888" value="0x002D4000" />
+        <register type="NOA" address="0x00009888" value="0x022D5000" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x102E0154" />
+        <register type="NOA" address="0x00009888" value="0x0C2E5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0055" />
+        <register type="NOA" address="0x00009888" value="0x104C8000" />
+        <register type="NOA" address="0x00009888" value="0x124C8000" />
+        <register type="NOA" address="0x00009888" value="0x144C8000" />
+        <register type="NOA" address="0x00009888" value="0x164C2000" />
+        <register type="NOA" address="0x00009888" value="0x044C8000" />
+        <register type="NOA" address="0x00009888" value="0x064CC000" />
+        <register type="NOA" address="0x00009888" value="0x084CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C4000" />
+        <register type="NOA" address="0x00009888" value="0x0C4EA000" />
+        <register type="NOA" address="0x00009888" value="0x0E4EA000" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x024EA000" />
+        <register type="NOA" address="0x00009888" value="0x044EA000" />
+        <register type="NOA" address="0x00009888" value="0x064E2000" />
+        <register type="NOA" address="0x00009888" value="0x1C4F5500" />
+        <register type="NOA" address="0x00009888" value="0x1A4F1554" />
+        <register type="NOA" address="0x00009888" value="0x0A640024" />
+        <register type="NOA" address="0x00009888" value="0x10640000" />
+        <register type="NOA" address="0x00009888" value="0x04640000" />
+        <register type="NOA" address="0x00009888" value="0x0C650024" />
+        <register type="NOA" address="0x00009888" value="0x10650000" />
+        <register type="NOA" address="0x00009888" value="0x06650000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C5327" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5425" />
+        <register type="NOA" address="0x00009888" value="0x006C2A00" />
+        <register type="NOA" address="0x00009888" value="0x026C285B" />
+        <register type="NOA" address="0x00009888" value="0x046C005C" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0900" />
+        <register type="NOA" address="0x00009888" value="0x1C0F0AA0" />
+        <register type="NOA" address="0x00009888" value="0x180F4000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F02AA" />
+        <register type="NOA" address="0x00009888" value="0x1C2C5400" />
+        <register type="NOA" address="0x00009888" value="0x1E2C0001" />
+        <register type="NOA" address="0x00009888" value="0x1A2C5550" />
+        <register type="NOA" address="0x00009888" value="0x1993AA00" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x05938000" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900421" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900001" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900420" />
+        <register type="NOA" address="0x00009888" value="0x45900021" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x03 ULT"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x14640340" />
+        <register type="NOA" address="0x00009888" value="0x14650340" />
+        <register type="NOA" address="0x00009888" value="0x106C6800" />
+        <register type="NOA" address="0x00009888" value="0x126C001E" />
+        <register type="NOA" address="0x00009888" value="0x166C0010" />
+        <register type="NOA" address="0x00009888" value="0x0C2D5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D5000" />
+        <register type="NOA" address="0x00009888" value="0x002D4000" />
+        <register type="NOA" address="0x00009888" value="0x022D5000" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x102E0154" />
+        <register type="NOA" address="0x00009888" value="0x0C2E5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0055" />
+        <register type="NOA" address="0x00009888" value="0x104C8000" />
+        <register type="NOA" address="0x00009888" value="0x124C8000" />
+        <register type="NOA" address="0x00009888" value="0x144C8000" />
+        <register type="NOA" address="0x00009888" value="0x164C2000" />
+        <register type="NOA" address="0x00009888" value="0x044C8000" />
+        <register type="NOA" address="0x00009888" value="0x064CC000" />
+        <register type="NOA" address="0x00009888" value="0x084CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C4000" />
+        <register type="NOA" address="0x00009888" value="0x0C4EA000" />
+        <register type="NOA" address="0x00009888" value="0x0E4EA000" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x024EA000" />
+        <register type="NOA" address="0x00009888" value="0x044EA000" />
+        <register type="NOA" address="0x00009888" value="0x064E2000" />
+        <register type="NOA" address="0x00009888" value="0x1C4F5500" />
+        <register type="NOA" address="0x00009888" value="0x1A4F1554" />
+        <register type="NOA" address="0x00009888" value="0x04642400" />
+        <register type="NOA" address="0x00009888" value="0x22640000" />
+        <register type="NOA" address="0x00009888" value="0x1A640000" />
+        <register type="NOA" address="0x00009888" value="0x06650024" />
+        <register type="NOA" address="0x00009888" value="0x22650000" />
+        <register type="NOA" address="0x00009888" value="0x1C650000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C5327" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5425" />
+        <register type="NOA" address="0x00009888" value="0x006C2A00" />
+        <register type="NOA" address="0x00009888" value="0x026C285B" />
+        <register type="NOA" address="0x00009888" value="0x046C005C" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0900" />
+        <register type="NOA" address="0x00009888" value="0x1C0F0AA0" />
+        <register type="NOA" address="0x00009888" value="0x180F4000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F02AA" />
+        <register type="NOA" address="0x00009888" value="0x1C2C5400" />
+        <register type="NOA" address="0x00009888" value="0x1E2C0001" />
+        <register type="NOA" address="0x00009888" value="0x1A2C5550" />
+        <register type="NOA" address="0x00009888" value="0x1993AA00" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x05938000" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900421" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900001" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900420" />
+        <register type="NOA" address="0x00009888" value="0x45900021" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="46dc44ca-491c-4cc1-a951-e7b3e62bf02b"
+       chipset="BXT"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x102D7800" />
+        <register type="NOA" address="0x00009888" value="0x122D79E0" />
+        <register type="NOA" address="0x00009888" value="0x0C2F0004" />
+        <register type="NOA" address="0x00009888" value="0x100E3800" />
+        <register type="NOA" address="0x00009888" value="0x180F0005" />
+        <register type="NOA" address="0x00009888" value="0x002D0940" />
+        <register type="NOA" address="0x00009888" value="0x022D802F" />
+        <register type="NOA" address="0x00009888" value="0x042D4013" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0050" />
+        <register type="NOA" address="0x00009888" value="0x022F0010" />
+        <register type="NOA" address="0x00009888" value="0x002F0000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C4000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E2000" />
+        <register type="NOA" address="0x00009888" value="0x040E0480" />
+        <register type="NOA" address="0x00009888" value="0x000E0000" />
+        <register type="NOA" address="0x00009888" value="0x060F0027" />
+        <register type="NOA" address="0x00009888" value="0x100F0000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F0040" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x05938000" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x439014A0" />
+        <register type="NOA" address="0x00009888" value="0x459000A4" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900001" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000EFFF" />
+        <register type="OA" address="0x00002778" value="0x00006000" />
+        <register type="OA" address="0x0000277C" value="0x0000F3FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler"
+       hw_config_guid="8364e2a8-af63-40af-b0d5-42969a255654"
+       chipset="BXT"
+       symbol_name="Sampler"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121300A0" />
+        <register type="NOA" address="0x00009888" value="0x141600AB" />
+        <register type="NOA" address="0x00009888" value="0x123300A0" />
+        <register type="NOA" address="0x00009888" value="0x143600AB" />
+        <register type="NOA" address="0x00009888" value="0x125300A0" />
+        <register type="NOA" address="0x00009888" value="0x145600AB" />
+        <register type="NOA" address="0x00009888" value="0x0C2D4000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D5000" />
+        <register type="NOA" address="0x00009888" value="0x002D4000" />
+        <register type="NOA" address="0x00009888" value="0x022D5000" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x102E01A0" />
+        <register type="NOA" address="0x00009888" value="0x0C2E5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0065" />
+        <register type="NOA" address="0x00009888" value="0x164C2000" />
+        <register type="NOA" address="0x00009888" value="0x044C8000" />
+        <register type="NOA" address="0x00009888" value="0x064CC000" />
+        <register type="NOA" address="0x00009888" value="0x084C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C4000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x024EA000" />
+        <register type="NOA" address="0x00009888" value="0x044E2000" />
+        <register type="NOA" address="0x00009888" value="0x064E2000" />
+        <register type="NOA" address="0x00009888" value="0x1C0F0800" />
+        <register type="NOA" address="0x00009888" value="0x180F4000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F023F" />
+        <register type="NOA" address="0x00009888" value="0x1E2C0003" />
+        <register type="NOA" address="0x00009888" value="0x1A2CC030" />
+        <register type="NOA" address="0x00009888" value="0x04132180" />
+        <register type="NOA" address="0x00009888" value="0x02130000" />
+        <register type="NOA" address="0x00009888" value="0x0C148000" />
+        <register type="NOA" address="0x00009888" value="0x0E142000" />
+        <register type="NOA" address="0x00009888" value="0x04148000" />
+        <register type="NOA" address="0x00009888" value="0x1E150140" />
+        <register type="NOA" address="0x00009888" value="0x1C150040" />
+        <register type="NOA" address="0x00009888" value="0x0C163000" />
+        <register type="NOA" address="0x00009888" value="0x0E160068" />
+        <register type="NOA" address="0x00009888" value="0x10160000" />
+        <register type="NOA" address="0x00009888" value="0x18160000" />
+        <register type="NOA" address="0x00009888" value="0x0A164000" />
+        <register type="NOA" address="0x00009888" value="0x04330043" />
+        <register type="NOA" address="0x00009888" value="0x02330000" />
+        <register type="NOA" address="0x00009888" value="0x0234A000" />
+        <register type="NOA" address="0x00009888" value="0x04342000" />
+        <register type="NOA" address="0x00009888" value="0x1C350015" />
+        <register type="NOA" address="0x00009888" value="0x02363460" />
+        <register type="NOA" address="0x00009888" value="0x10360000" />
+        <register type="NOA" address="0x00009888" value="0x04360000" />
+        <register type="NOA" address="0x00009888" value="0x06360000" />
+        <register type="NOA" address="0x00009888" value="0x08364000" />
+        <register type="NOA" address="0x00009888" value="0x06530043" />
+        <register type="NOA" address="0x00009888" value="0x02530000" />
+        <register type="NOA" address="0x00009888" value="0x0E548000" />
+        <register type="NOA" address="0x00009888" value="0x00548000" />
+        <register type="NOA" address="0x00009888" value="0x06542000" />
+        <register type="NOA" address="0x00009888" value="0x1E550400" />
+        <register type="NOA" address="0x00009888" value="0x1A552000" />
+        <register type="NOA" address="0x00009888" value="0x1C550100" />
+        <register type="NOA" address="0x00009888" value="0x0E563000" />
+        <register type="NOA" address="0x00009888" value="0x00563400" />
+        <register type="NOA" address="0x00009888" value="0x10560000" />
+        <register type="NOA" address="0x00009888" value="0x18560000" />
+        <register type="NOA" address="0x00009888" value="0x02560000" />
+        <register type="NOA" address="0x00009888" value="0x0C564000" />
+        <register type="NOA" address="0x00009888" value="0x1993A800" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x05938000" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B9014A0" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900001" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900820" />
+        <register type="NOA" address="0x00009888" value="0x45901022" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="175c8092-cb25-4d1e-8dc7-b4fdd39e2d92"
+       chipset="BXT"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A0000" />
+        <register type="NOA" address="0x00009888" value="0x143A0000" />
+        <register type="NOA" address="0x00009888" value="0x145A0000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D4000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D5000" />
+        <register type="NOA" address="0x00009888" value="0x002D4000" />
+        <register type="NOA" address="0x00009888" value="0x022D5000" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x102E0150" />
+        <register type="NOA" address="0x00009888" value="0x0C2E5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E006A" />
+        <register type="NOA" address="0x00009888" value="0x124C8000" />
+        <register type="NOA" address="0x00009888" value="0x144C8000" />
+        <register type="NOA" address="0x00009888" value="0x164C2000" />
+        <register type="NOA" address="0x00009888" value="0x044C8000" />
+        <register type="NOA" address="0x00009888" value="0x064C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C4000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4EA000" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x024E2000" />
+        <register type="NOA" address="0x00009888" value="0x064E2000" />
+        <register type="NOA" address="0x00009888" value="0x1C0F0BC0" />
+        <register type="NOA" address="0x00009888" value="0x180F4000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F0302" />
+        <register type="NOA" address="0x00009888" value="0x1E2C0003" />
+        <register type="NOA" address="0x00009888" value="0x1A2C00F0" />
+        <register type="NOA" address="0x00009888" value="0x021A3080" />
+        <register type="NOA" address="0x00009888" value="0x041A31E5" />
+        <register type="NOA" address="0x00009888" value="0x02148000" />
+        <register type="NOA" address="0x00009888" value="0x0414A000" />
+        <register type="NOA" address="0x00009888" value="0x1C150054" />
+        <register type="NOA" address="0x00009888" value="0x06168000" />
+        <register type="NOA" address="0x00009888" value="0x08168000" />
+        <register type="NOA" address="0x00009888" value="0x0A168000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A3280" />
+        <register type="NOA" address="0x00009888" value="0x0E3A0063" />
+        <register type="NOA" address="0x00009888" value="0x063A0061" />
+        <register type="NOA" address="0x00009888" value="0x023A0000" />
+        <register type="NOA" address="0x00009888" value="0x0C348000" />
+        <register type="NOA" address="0x00009888" value="0x0E342000" />
+        <register type="NOA" address="0x00009888" value="0x06342000" />
+        <register type="NOA" address="0x00009888" value="0x1E350140" />
+        <register type="NOA" address="0x00009888" value="0x1C350100" />
+        <register type="NOA" address="0x00009888" value="0x18360028" />
+        <register type="NOA" address="0x00009888" value="0x0C368000" />
+        <register type="NOA" address="0x00009888" value="0x0E5A3080" />
+        <register type="NOA" address="0x00009888" value="0x005A3280" />
+        <register type="NOA" address="0x00009888" value="0x025A0063" />
+        <register type="NOA" address="0x00009888" value="0x0E548000" />
+        <register type="NOA" address="0x00009888" value="0x00548000" />
+        <register type="NOA" address="0x00009888" value="0x02542000" />
+        <register type="NOA" address="0x00009888" value="0x1E550400" />
+        <register type="NOA" address="0x00009888" value="0x1A552000" />
+        <register type="NOA" address="0x00009888" value="0x1C550001" />
+        <register type="NOA" address="0x00009888" value="0x18560080" />
+        <register type="NOA" address="0x00009888" value="0x02568000" />
+        <register type="NOA" address="0x00009888" value="0x04568000" />
+        <register type="NOA" address="0x00009888" value="0x1993A800" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x05938000" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x2D904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900420" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4D900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x45901084" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900001" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x00007FFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x00009FFF" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000EFFF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000F3FF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FDFF" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="d260f03f-b34d-4b49-a44e-436819117332"
+       chipset="BXT"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A026B" />
+        <register type="NOA" address="0x00009888" value="0x143A0173" />
+        <register type="NOA" address="0x00009888" value="0x145A026B" />
+        <register type="NOA" address="0x00009888" value="0x002D4000" />
+        <register type="NOA" address="0x00009888" value="0x022D5000" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x0C2E5000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0069" />
+        <register type="NOA" address="0x00009888" value="0x044C8000" />
+        <register type="NOA" address="0x00009888" value="0x064CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C4000" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x024EA000" />
+        <register type="NOA" address="0x00009888" value="0x064E2000" />
+        <register type="NOA" address="0x00009888" value="0x180F6000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F030A" />
+        <register type="NOA" address="0x00009888" value="0x1A2C03C0" />
+        <register type="NOA" address="0x00009888" value="0x041A37E7" />
+        <register type="NOA" address="0x00009888" value="0x021A0000" />
+        <register type="NOA" address="0x00009888" value="0x0414A000" />
+        <register type="NOA" address="0x00009888" value="0x1C150050" />
+        <register type="NOA" address="0x00009888" value="0x08168000" />
+        <register type="NOA" address="0x00009888" value="0x0A168000" />
+        <register type="NOA" address="0x00009888" value="0x003A3380" />
+        <register type="NOA" address="0x00009888" value="0x063A006F" />
+        <register type="NOA" address="0x00009888" value="0x023A0000" />
+        <register type="NOA" address="0x00009888" value="0x00348000" />
+        <register type="NOA" address="0x00009888" value="0x06342000" />
+        <register type="NOA" address="0x00009888" value="0x1A352000" />
+        <register type="NOA" address="0x00009888" value="0x1C350100" />
+        <register type="NOA" address="0x00009888" value="0x02368000" />
+        <register type="NOA" address="0x00009888" value="0x0C368000" />
+        <register type="NOA" address="0x00009888" value="0x025A37E7" />
+        <register type="NOA" address="0x00009888" value="0x0254A000" />
+        <register type="NOA" address="0x00009888" value="0x1C550005" />
+        <register type="NOA" address="0x00009888" value="0x04568000" />
+        <register type="NOA" address="0x00009888" value="0x06568000" />
+        <register type="NOA" address="0x00009888" value="0x03938000" />
+        <register type="NOA" address="0x00009888" value="0x05938000" />
+        <register type="NOA" address="0x00009888" value="0x07938000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17904000" />
+        <register type="NOA" address="0x00009888" value="0x19904000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900020" />
+        <register type="NOA" address="0x00009888" value="0x45901080" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900001" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extra Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extra"
+       hw_config_guid="fa6ecf21-2cb8-4d0b-9308-6e4a7b4ca87a"
+       chipset="BXT"
+       symbol_name="ComputeExtra"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active including Ext Math"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing including Extended Math processing"
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ  C 7 READ C 6 READ FADD C 5 READ FADD 8 FMUL FADD 100 FMUL $EuCoresTotalCount FDIV $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active_adjusted"
+             units="percent"
+             symbol_name="Fpu1ActiveAdjusted"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A001F" />
+        <register type="NOA" address="0x00009888" value="0x143A001F" />
+        <register type="NOA" address="0x00009888" value="0x145A001F" />
+        <register type="NOA" address="0x00009888" value="0x042D5000" />
+        <register type="NOA" address="0x00009888" value="0x062D1000" />
+        <register type="NOA" address="0x00009888" value="0x0E2E0094" />
+        <register type="NOA" address="0x00009888" value="0x084CC000" />
+        <register type="NOA" address="0x00009888" value="0x044EA000" />
+        <register type="NOA" address="0x00009888" value="0x1A0F00E0" />
+        <register type="NOA" address="0x00009888" value="0x1A2C0C00" />
+        <register type="NOA" address="0x00009888" value="0x061A0063" />
+        <register type="NOA" address="0x00009888" value="0x021A0000" />
+        <register type="NOA" address="0x00009888" value="0x06142000" />
+        <register type="NOA" address="0x00009888" value="0x1C150100" />
+        <register type="NOA" address="0x00009888" value="0x0C168000" />
+        <register type="NOA" address="0x00009888" value="0x043A3180" />
+        <register type="NOA" address="0x00009888" value="0x023A0000" />
+        <register type="NOA" address="0x00009888" value="0x04348000" />
+        <register type="NOA" address="0x00009888" value="0x1C350040" />
+        <register type="NOA" address="0x00009888" value="0x0A368000" />
+        <register type="NOA" address="0x00009888" value="0x045A0063" />
+        <register type="NOA" address="0x00009888" value="0x025A0000" />
+        <register type="NOA" address="0x00009888" value="0x04542000" />
+        <register type="NOA" address="0x00009888" value="0x1C550010" />
+        <register type="NOA" address="0x00009888" value="0x08568000" />
+        <register type="NOA" address="0x00009888" value="0x09938000" />
+        <register type="NOA" address="0x00009888" value="0x0B938000" />
+        <register type="NOA" address="0x00009888" value="0x0D938000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1D904000" />
+        <register type="NOA" address="0x00009888" value="0x1F904000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+        <register type="NOA" address="0x00009888" value="0x47900004" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00001000" />
+        <register type="FLEX" address="0x0000E558" value="0x00003002" />
+        <register type="FLEX" address="0x0000E658" value="0x00005004" />
+        <register type="FLEX" address="0x0000E758" value="0x00011010" />
+        <register type="FLEX" address="0x0000E45C" value="0x00050012" />
+        <register type="FLEX" address="0x0000E55C" value="0x00052051" />
+        <register type="FLEX" address="0x0000E65C" value="0x00000008" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="5ee72f5c-092f-421e-8b70-225f7c3e9612"
+       chipset="BXT"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x19800000" />
+        <register type="NOA" address="0x00009888" value="0x07800063" />
+        <register type="NOA" address="0x00009888" value="0x11800000" />
+        <register type="NOA" address="0x00009888" value="0x23810008" />
+        <register type="NOA" address="0x00009888" value="0x1D950400" />
+        <register type="NOA" address="0x00009888" value="0x0F922000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
diff --git a/src/mesa/drivers/dri/i965/brw_oa_chv.xml b/src/mesa/drivers/dri/i965/brw_oa_chv.xml
new file mode 100644
index 0000000..a4ebfb3
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_chv.xml
@@ -0,0 +1,9569 @@
+<?xml version="1.0"?>
+<metrics version="1491577975" merge_md5="">
+  <set name="Render Metrics Basic Gen8LP"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="9d8a3af5-c02c-4a4a-b947-f1672469e0fb"
+       chipset="CHV"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 1 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="A 29 READ 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ  B 1 READ UADD $GpuCoreClocks FDIV 2 FDIV 100 FMUL"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+        <register type="NOA" address="0x00009888" value="0x59800001" />
+        <register type="NOA" address="0x00009888" value="0x285A0006" />
+        <register type="NOA" address="0x00009888" value="0x2C110014" />
+        <register type="NOA" address="0x00009888" value="0x2E110000" />
+        <register type="NOA" address="0x00009888" value="0x2C310014" />
+        <register type="NOA" address="0x00009888" value="0x2E310000" />
+        <register type="NOA" address="0x00009888" value="0x2B8303DF" />
+        <register type="NOA" address="0x00009888" value="0x3580024F" />
+        <register type="NOA" address="0x00009888" value="0x00580888" />
+        <register type="NOA" address="0x00009888" value="0x1E5A0015" />
+        <register type="NOA" address="0x00009888" value="0x205A0014" />
+        <register type="NOA" address="0x00009888" value="0x045A0000" />
+        <register type="NOA" address="0x00009888" value="0x025A0000" />
+        <register type="NOA" address="0x00009888" value="0x02180500" />
+        <register type="NOA" address="0x00009888" value="0x00190555" />
+        <register type="NOA" address="0x00009888" value="0x021D0500" />
+        <register type="NOA" address="0x00009888" value="0x021F0A00" />
+        <register type="NOA" address="0x00009888" value="0x00380444" />
+        <register type="NOA" address="0x00009888" value="0x02390500" />
+        <register type="NOA" address="0x00009888" value="0x003A0666" />
+        <register type="NOA" address="0x00009888" value="0x00100111" />
+        <register type="NOA" address="0x00009888" value="0x06110030" />
+        <register type="NOA" address="0x00009888" value="0x0A110031" />
+        <register type="NOA" address="0x00009888" value="0x0E110046" />
+        <register type="NOA" address="0x00009888" value="0x04110000" />
+        <register type="NOA" address="0x00009888" value="0x00110000" />
+        <register type="NOA" address="0x00009888" value="0x00130111" />
+        <register type="NOA" address="0x00009888" value="0x00300444" />
+        <register type="NOA" address="0x00009888" value="0x08310030" />
+        <register type="NOA" address="0x00009888" value="0x0C310031" />
+        <register type="NOA" address="0x00009888" value="0x10310046" />
+        <register type="NOA" address="0x00009888" value="0x04310000" />
+        <register type="NOA" address="0x00009888" value="0x00310000" />
+        <register type="NOA" address="0x00009888" value="0x00330444" />
+        <register type="NOA" address="0x00009888" value="0x038A0A00" />
+        <register type="NOA" address="0x00009888" value="0x018B0FFF" />
+        <register type="NOA" address="0x00009888" value="0x038B0A00" />
+        <register type="NOA" address="0x00009888" value="0x01855000" />
+        <register type="NOA" address="0x00009888" value="0x03850055" />
+        <register type="NOA" address="0x00009888" value="0x13830021" />
+        <register type="NOA" address="0x00009888" value="0x15830020" />
+        <register type="NOA" address="0x00009888" value="0x1783002F" />
+        <register type="NOA" address="0x00009888" value="0x1983002E" />
+        <register type="NOA" address="0x00009888" value="0x1B83002D" />
+        <register type="NOA" address="0x00009888" value="0x1D83002C" />
+        <register type="NOA" address="0x00009888" value="0x05830000" />
+        <register type="NOA" address="0x00009888" value="0x01840555" />
+        <register type="NOA" address="0x00009888" value="0x03840500" />
+        <register type="NOA" address="0x00009888" value="0x23800074" />
+        <register type="NOA" address="0x00009888" value="0x2580007D" />
+        <register type="NOA" address="0x00009888" value="0x05800000" />
+        <register type="NOA" address="0x00009888" value="0x01805000" />
+        <register type="NOA" address="0x00009888" value="0x03800055" />
+        <register type="NOA" address="0x00009888" value="0x01865000" />
+        <register type="NOA" address="0x00009888" value="0x03860055" />
+        <register type="NOA" address="0x00009888" value="0x01875000" />
+        <register type="NOA" address="0x00009888" value="0x03870055" />
+        <register type="NOA" address="0x00009888" value="0x418000AA" />
+        <register type="NOA" address="0x00009888" value="0x4380000A" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x4780000A" />
+        <register type="NOA" address="0x00009888" value="0x49800000" />
+        <register type="NOA" address="0x00009888" value="0x4B800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x55800000" />
+        <register type="NOA" address="0x00009888" value="0x57800000" />
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen8LP"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="f522a89c-ecd1-4522-8331-3383c54af5f5"
+       chipset="CHV"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 4 READ B 5 READ UADD 64 UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 2 READ B 3 READ UADD 64 UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 0 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Ring Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and Uncore ring."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 128 UMUL"
+             equation="C 2 READ 128 UMUL"
+             underscore_name="gti_ring_throughput"
+             units="bytes"
+             symbol_name="GtiRingThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 0 READ B 1 READ UADD 64 UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read-Only Stall"
+             description="The percentage of time in which GTI Read-Only port has been stalled."
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gti_ro_stall"
+             units="percent"
+             symbol_name="GtiRoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read-Write Stall"
+             description="The percentage of time in which GTI Read-Write port has been stalled."
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gti_rw_stall"
+             units="percent"
+             symbol_name="GtiRwStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 1 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+        <register type="NOA" address="0x00009888" value="0x59800001" />
+        <register type="NOA" address="0x00009888" value="0x2E5800E0" />
+        <register type="NOA" address="0x00009888" value="0x2E3800E0" />
+        <register type="NOA" address="0x00009888" value="0x3580024F" />
+        <register type="NOA" address="0x00009888" value="0x3D800140" />
+        <register type="NOA" address="0x00009888" value="0x08580042" />
+        <register type="NOA" address="0x00009888" value="0x0C580040" />
+        <register type="NOA" address="0x00009888" value="0x1058004C" />
+        <register type="NOA" address="0x00009888" value="0x1458004B" />
+        <register type="NOA" address="0x00009888" value="0x04580000" />
+        <register type="NOA" address="0x00009888" value="0x00580000" />
+        <register type="NOA" address="0x00009888" value="0x00195555" />
+        <register type="NOA" address="0x00009888" value="0x06380042" />
+        <register type="NOA" address="0x00009888" value="0x0A380040" />
+        <register type="NOA" address="0x00009888" value="0x0E38004C" />
+        <register type="NOA" address="0x00009888" value="0x1238004B" />
+        <register type="NOA" address="0x00009888" value="0x04380000" />
+        <register type="NOA" address="0x00009888" value="0x00384444" />
+        <register type="NOA" address="0x00009888" value="0x003A5555" />
+        <register type="NOA" address="0x00009888" value="0x018BFFFF" />
+        <register type="NOA" address="0x00009888" value="0x01845555" />
+        <register type="NOA" address="0x00009888" value="0x17800074" />
+        <register type="NOA" address="0x00009888" value="0x1980007D" />
+        <register type="NOA" address="0x00009888" value="0x1B80007C" />
+        <register type="NOA" address="0x00009888" value="0x1D8000B6" />
+        <register type="NOA" address="0x00009888" value="0x1F8000B7" />
+        <register type="NOA" address="0x00009888" value="0x05800000" />
+        <register type="NOA" address="0x00009888" value="0x03800000" />
+        <register type="NOA" address="0x00009888" value="0x418000AA" />
+        <register type="NOA" address="0x00009888" value="0x438000AA" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x47800000" />
+        <register type="NOA" address="0x00009888" value="0x4980012A" />
+        <register type="NOA" address="0x00009888" value="0x4B80012A" />
+        <register type="NOA" address="0x00009888" value="0x4D80012A" />
+        <register type="NOA" address="0x00009888" value="0x4F80012A" />
+        <register type="NOA" address="0x00009888" value="0x518001CE" />
+        <register type="NOA" address="0x00009888" value="0x538001CE" />
+        <register type="NOA" address="0x00009888" value="0x5580000E" />
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="a9ccc03d-a943-4e6b-9cd6-13e063075927"
+       chipset="CHV"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses UADD UADD 64 UMUL"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+        <register type="NOA" address="0x00009888" value="0x59800001" />
+        <register type="NOA" address="0x00009888" value="0x261E0000" />
+        <register type="NOA" address="0x00009888" value="0x281F000F" />
+        <register type="NOA" address="0x00009888" value="0x2817001A" />
+        <register type="NOA" address="0x00009888" value="0x2791001F" />
+        <register type="NOA" address="0x00009888" value="0x27880019" />
+        <register type="NOA" address="0x00009888" value="0x2D890000" />
+        <register type="NOA" address="0x00009888" value="0x278A0007" />
+        <register type="NOA" address="0x00009888" value="0x298D001F" />
+        <register type="NOA" address="0x00009888" value="0x278E0020" />
+        <register type="NOA" address="0x00009888" value="0x2B8F0012" />
+        <register type="NOA" address="0x00009888" value="0x29900000" />
+        <register type="NOA" address="0x00009888" value="0x00184000" />
+        <register type="NOA" address="0x00009888" value="0x02181000" />
+        <register type="NOA" address="0x00009888" value="0x02194000" />
+        <register type="NOA" address="0x00009888" value="0x141E0002" />
+        <register type="NOA" address="0x00009888" value="0x041E0000" />
+        <register type="NOA" address="0x00009888" value="0x001E0000" />
+        <register type="NOA" address="0x00009888" value="0x221F0015" />
+        <register type="NOA" address="0x00009888" value="0x041F0000" />
+        <register type="NOA" address="0x00009888" value="0x001F4000" />
+        <register type="NOA" address="0x00009888" value="0x021F0000" />
+        <register type="NOA" address="0x00009888" value="0x023A8000" />
+        <register type="NOA" address="0x00009888" value="0x0213C000" />
+        <register type="NOA" address="0x00009888" value="0x02164000" />
+        <register type="NOA" address="0x00009888" value="0x24170012" />
+        <register type="NOA" address="0x00009888" value="0x04170000" />
+        <register type="NOA" address="0x00009888" value="0x07910005" />
+        <register type="NOA" address="0x00009888" value="0x05910000" />
+        <register type="NOA" address="0x00009888" value="0x01911500" />
+        <register type="NOA" address="0x00009888" value="0x03910501" />
+        <register type="NOA" address="0x00009888" value="0x0D880002" />
+        <register type="NOA" address="0x00009888" value="0x1D880003" />
+        <register type="NOA" address="0x00009888" value="0x05880000" />
+        <register type="NOA" address="0x00009888" value="0x0B890032" />
+        <register type="NOA" address="0x00009888" value="0x1B890031" />
+        <register type="NOA" address="0x00009888" value="0x05890000" />
+        <register type="NOA" address="0x00009888" value="0x01890040" />
+        <register type="NOA" address="0x00009888" value="0x03890040" />
+        <register type="NOA" address="0x00009888" value="0x098A0000" />
+        <register type="NOA" address="0x00009888" value="0x198A0004" />
+        <register type="NOA" address="0x00009888" value="0x058A0000" />
+        <register type="NOA" address="0x00009888" value="0x018A8050" />
+        <register type="NOA" address="0x00009888" value="0x038A2050" />
+        <register type="NOA" address="0x00009888" value="0x018B95A9" />
+        <register type="NOA" address="0x00009888" value="0x038BE5A9" />
+        <register type="NOA" address="0x00009888" value="0x018C1500" />
+        <register type="NOA" address="0x00009888" value="0x038C0501" />
+        <register type="NOA" address="0x00009888" value="0x178D0015" />
+        <register type="NOA" address="0x00009888" value="0x058D0000" />
+        <register type="NOA" address="0x00009888" value="0x138E0004" />
+        <register type="NOA" address="0x00009888" value="0x218E000C" />
+        <register type="NOA" address="0x00009888" value="0x058E0000" />
+        <register type="NOA" address="0x00009888" value="0x018E0500" />
+        <register type="NOA" address="0x00009888" value="0x038E0101" />
+        <register type="NOA" address="0x00009888" value="0x0F8F0027" />
+        <register type="NOA" address="0x00009888" value="0x058F0000" />
+        <register type="NOA" address="0x00009888" value="0x018F0000" />
+        <register type="NOA" address="0x00009888" value="0x038F0001" />
+        <register type="NOA" address="0x00009888" value="0x11900013" />
+        <register type="NOA" address="0x00009888" value="0x1F900017" />
+        <register type="NOA" address="0x00009888" value="0x05900000" />
+        <register type="NOA" address="0x00009888" value="0x01900100" />
+        <register type="NOA" address="0x00009888" value="0x03900001" />
+        <register type="NOA" address="0x00009888" value="0x01845555" />
+        <register type="NOA" address="0x00009888" value="0x03845555" />
+        <register type="NOA" address="0x00009888" value="0x418000AA" />
+        <register type="NOA" address="0x00009888" value="0x438000AA" />
+        <register type="NOA" address="0x00009888" value="0x458000AA" />
+        <register type="NOA" address="0x00009888" value="0x478000AA" />
+        <register type="NOA" address="0x00009888" value="0x4980018C" />
+        <register type="NOA" address="0x00009888" value="0x4B80014B" />
+        <register type="NOA" address="0x00009888" value="0x4D800128" />
+        <register type="NOA" address="0x00009888" value="0x4F80012A" />
+        <register type="NOA" address="0x00009888" value="0x51800187" />
+        <register type="NOA" address="0x00009888" value="0x5380014B" />
+        <register type="NOA" address="0x00009888" value="0x55800149" />
+        <register type="NOA" address="0x00009888" value="0x5780010A" />
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="2cf0c064-68df-4fac-9b3f-57f51ca8a069"
+       chipset="CHV"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 1 READ C 0 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader12_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader12AccessStalledOnL3"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader02_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader02AccessStalledOnL3"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ B 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader10_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader10AccessStalledOnL3"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="B 7 READ B 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader11_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader11AccessStalledOnL3"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ C 2 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x105C0232" />
+        <register type="NOA" address="0x00009888" value="0x10580232" />
+        <register type="NOA" address="0x00009888" value="0x10380232" />
+        <register type="NOA" address="0x00009888" value="0x10DC0232" />
+        <register type="NOA" address="0x00009888" value="0x10D80232" />
+        <register type="NOA" address="0x00009888" value="0x10B80232" />
+        <register type="NOA" address="0x00009888" value="0x118E4400" />
+        <register type="NOA" address="0x00009888" value="0x025C6080" />
+        <register type="NOA" address="0x00009888" value="0x045C004B" />
+        <register type="NOA" address="0x00009888" value="0x005C8000" />
+        <register type="NOA" address="0x00009888" value="0x00582080" />
+        <register type="NOA" address="0x00009888" value="0x0258004B" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00AA" />
+        <register type="NOA" address="0x00009888" value="0x04386080" />
+        <register type="NOA" address="0x00009888" value="0x0638404B" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A380000" />
+        <register type="NOA" address="0x00009888" value="0x0C380000" />
+        <register type="NOA" address="0x00009888" value="0x00398000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x0CDC25C1" />
+        <register type="NOA" address="0x00009888" value="0x0ADCC000" />
+        <register type="NOA" address="0x00009888" value="0x0AD825C1" />
+        <register type="NOA" address="0x00009888" value="0x18DB4000" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0001" />
+        <register type="NOA" address="0x00009888" value="0x0E9F8000" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0EB825C1" />
+        <register type="NOA" address="0x00009888" value="0x18B80154" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0D88C000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258BAA05" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x198C5400" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x098DC000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x098E05C0" />
+        <register type="NOA" address="0x00009888" value="0x058E0000" />
+        <register type="NOA" address="0x00009888" value="0x198F0020" />
+        <register type="NOA" address="0x00009888" value="0x2185AA0A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x19835000" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x19808000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x51800040" />
+        <register type="NOA" address="0x00009888" value="0x43800400" />
+        <register type="NOA" address="0x00009888" value="0x45800800" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800C62" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F801042" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x418014A4" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FFF7" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="78a87ff9-543a-49ce-95ea-26d86071ea93"
+       chipset="CHV"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank1 Stalled"
+             description="The percentage of time in which slice1 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank1_stalled"
+             units="percent"
+             symbol_name="L31Bank1Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice1 L3 Bank0 Stalled"
+             description="The percentage of time in which slice1 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank0_stalled"
+             units="percent"
+             symbol_name="L31Bank0Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Slice1 L3 Bank1 Active"
+             description="The percentage of time in which slice1 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank1_active"
+             units="percent"
+             symbol_name="L31Bank1Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank0 Active"
+             description="The percentage of time in which slice1 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank0_active"
+             units="percent"
+             symbol_name="L31Bank0Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x10BF03DA" />
+        <register type="NOA" address="0x00009888" value="0x14BF0001" />
+        <register type="NOA" address="0x00009888" value="0x12980340" />
+        <register type="NOA" address="0x00009888" value="0x12990340" />
+        <register type="NOA" address="0x00009888" value="0x0CBF1187" />
+        <register type="NOA" address="0x00009888" value="0x0EBF1205" />
+        <register type="NOA" address="0x00009888" value="0x00BF0500" />
+        <register type="NOA" address="0x00009888" value="0x02BF042B" />
+        <register type="NOA" address="0x00009888" value="0x04BF002C" />
+        <register type="NOA" address="0x00009888" value="0x0CDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDAC000" />
+        <register type="NOA" address="0x00009888" value="0x00DA8000" />
+        <register type="NOA" address="0x00009888" value="0x02DAC000" />
+        <register type="NOA" address="0x00009888" value="0x04DA4000" />
+        <register type="NOA" address="0x00009888" value="0x04983400" />
+        <register type="NOA" address="0x00009888" value="0x10980000" />
+        <register type="NOA" address="0x00009888" value="0x06990034" />
+        <register type="NOA" address="0x00009888" value="0x10990000" />
+        <register type="NOA" address="0x00009888" value="0x0C9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E9DC000" />
+        <register type="NOA" address="0x00009888" value="0x009D8000" />
+        <register type="NOA" address="0x00009888" value="0x029DC000" />
+        <register type="NOA" address="0x00009888" value="0x049D4000" />
+        <register type="NOA" address="0x00009888" value="0x109F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F00BA" />
+        <register type="NOA" address="0x00009888" value="0x0CB88000" />
+        <register type="NOA" address="0x00009888" value="0x0CB95000" />
+        <register type="NOA" address="0x00009888" value="0x0EB95000" />
+        <register type="NOA" address="0x00009888" value="0x00B94000" />
+        <register type="NOA" address="0x00009888" value="0x02B95000" />
+        <register type="NOA" address="0x00009888" value="0x04B91000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x0CBA4000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x258B800A" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B5500" />
+        <register type="NOA" address="0x00009888" value="0x198C4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x47800000" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_2"
+       hw_config_guid="9f2cece5-7bfe-4320-ad66-8c7cc526bec5"
+       chipset="CHV"
+       symbol_name="L3_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x103F03DA" />
+        <register type="NOA" address="0x00009888" value="0x143F0001" />
+        <register type="NOA" address="0x00009888" value="0x12180340" />
+        <register type="NOA" address="0x00009888" value="0x12190340" />
+        <register type="NOA" address="0x00009888" value="0x0C3F1187" />
+        <register type="NOA" address="0x00009888" value="0x0E3F1205" />
+        <register type="NOA" address="0x00009888" value="0x003F0500" />
+        <register type="NOA" address="0x00009888" value="0x023F042B" />
+        <register type="NOA" address="0x00009888" value="0x043F002C" />
+        <register type="NOA" address="0x00009888" value="0x0C5AC000" />
+        <register type="NOA" address="0x00009888" value="0x0E5AC000" />
+        <register type="NOA" address="0x00009888" value="0x005A8000" />
+        <register type="NOA" address="0x00009888" value="0x025AC000" />
+        <register type="NOA" address="0x00009888" value="0x045A4000" />
+        <register type="NOA" address="0x00009888" value="0x04183400" />
+        <register type="NOA" address="0x00009888" value="0x10180000" />
+        <register type="NOA" address="0x00009888" value="0x06190034" />
+        <register type="NOA" address="0x00009888" value="0x10190000" />
+        <register type="NOA" address="0x00009888" value="0x0C1DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1DC000" />
+        <register type="NOA" address="0x00009888" value="0x001D8000" />
+        <register type="NOA" address="0x00009888" value="0x021DC000" />
+        <register type="NOA" address="0x00009888" value="0x041D4000" />
+        <register type="NOA" address="0x00009888" value="0x101F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00BA" />
+        <register type="NOA" address="0x00009888" value="0x0C388000" />
+        <register type="NOA" address="0x00009888" value="0x0C395000" />
+        <register type="NOA" address="0x00009888" value="0x0E395000" />
+        <register type="NOA" address="0x00009888" value="0x00394000" />
+        <register type="NOA" address="0x00009888" value="0x02395000" />
+        <register type="NOA" address="0x00009888" value="0x04391000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AA800" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258B4005" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800000" />
+        <register type="NOA" address="0x00009888" value="0x47800000" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_3"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_3"
+       hw_config_guid="d890ef38-d309-47e4-b8b5-aa779bb19ab0"
+       chipset="CHV"
+       symbol_name="L3_3"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 L3 Bank3 Stalled"
+             description="The percentage of time in which slice0 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_stalled"
+             units="percent"
+             symbol_name="L30Bank3Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank3 Active"
+             description="The percentage of time in which slice0 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_active"
+             units="percent"
+             symbol_name="L30Bank3Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank3 Active"
+             description="The percentage of time in which slice1 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank3_active"
+             units="percent"
+             symbol_name="L31Bank3Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice1 L3 Bank3 Stalled"
+             description="The percentage of time in which slice1 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank3_stalled"
+             units="percent"
+             symbol_name="L31Bank3Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121B0340" />
+        <register type="NOA" address="0x00009888" value="0x103F0274" />
+        <register type="NOA" address="0x00009888" value="0x123F0000" />
+        <register type="NOA" address="0x00009888" value="0x129B0340" />
+        <register type="NOA" address="0x00009888" value="0x10BF0274" />
+        <register type="NOA" address="0x00009888" value="0x12BF0000" />
+        <register type="NOA" address="0x00009888" value="0x041B3400" />
+        <register type="NOA" address="0x00009888" value="0x101B0000" />
+        <register type="NOA" address="0x00009888" value="0x045C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A3D4000" />
+        <register type="NOA" address="0x00009888" value="0x003F0080" />
+        <register type="NOA" address="0x00009888" value="0x023F0793" />
+        <register type="NOA" address="0x00009888" value="0x043F0014" />
+        <register type="NOA" address="0x00009888" value="0x04588000" />
+        <register type="NOA" address="0x00009888" value="0x005A8000" />
+        <register type="NOA" address="0x00009888" value="0x025AC000" />
+        <register type="NOA" address="0x00009888" value="0x045A4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x001D8000" />
+        <register type="NOA" address="0x00009888" value="0x021DC000" />
+        <register type="NOA" address="0x00009888" value="0x041D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F002A" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x00394000" />
+        <register type="NOA" address="0x00009888" value="0x02395000" />
+        <register type="NOA" address="0x00009888" value="0x04399000" />
+        <register type="NOA" address="0x00009888" value="0x069B0034" />
+        <register type="NOA" address="0x00009888" value="0x109B0000" />
+        <register type="NOA" address="0x00009888" value="0x06DC4000" />
+        <register type="NOA" address="0x00009888" value="0x0CBD4000" />
+        <register type="NOA" address="0x00009888" value="0x0CBF0981" />
+        <register type="NOA" address="0x00009888" value="0x0EBF0A0F" />
+        <register type="NOA" address="0x00009888" value="0x06D84000" />
+        <register type="NOA" address="0x00009888" value="0x0CDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0CDB4000" />
+        <register type="NOA" address="0x00009888" value="0x0C9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E9DC000" />
+        <register type="NOA" address="0x00009888" value="0x109F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0E9F0080" />
+        <register type="NOA" address="0x00009888" value="0x0CB84000" />
+        <register type="NOA" address="0x00009888" value="0x0CB95000" />
+        <register type="NOA" address="0x00009888" value="0x0EB95000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258B8009" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x198C4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800C00" />
+        <register type="NOA" address="0x00009888" value="0x47800C63" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F8014A5" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800045" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_4"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_4"
+       hw_config_guid="5fdff4a6-9dc8-45e1-bfda-ef54869fbdd4"
+       chipset="CHV"
+       symbol_name="L3_4"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Stalled"
+             description="The percentage of time in which slice0 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_stalled"
+             units="percent"
+             symbol_name="L30Bank2Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 L3 Bank2 Active"
+             description="The percentage of time in which slice1 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank2_active"
+             units="percent"
+             symbol_name="L31Bank2Active"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 L3 Bank2 Active"
+             description="The percentage of time in which slice0 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_active"
+             units="percent"
+             symbol_name="L30Bank2Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Slice1 L3 Bank2 Stalled"
+             description="The percentage of time in which slice1 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l31_bank2_stalled"
+             units="percent"
+             symbol_name="L31Bank2Stalled"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121A0340" />
+        <register type="NOA" address="0x00009888" value="0x103F0017" />
+        <register type="NOA" address="0x00009888" value="0x123F0020" />
+        <register type="NOA" address="0x00009888" value="0x129A0340" />
+        <register type="NOA" address="0x00009888" value="0x10BF0017" />
+        <register type="NOA" address="0x00009888" value="0x12BF0020" />
+        <register type="NOA" address="0x00009888" value="0x041A3400" />
+        <register type="NOA" address="0x00009888" value="0x101A0000" />
+        <register type="NOA" address="0x00009888" value="0x043B8000" />
+        <register type="NOA" address="0x00009888" value="0x0A3E0010" />
+        <register type="NOA" address="0x00009888" value="0x003F0200" />
+        <register type="NOA" address="0x00009888" value="0x023F0113" />
+        <register type="NOA" address="0x00009888" value="0x043F0014" />
+        <register type="NOA" address="0x00009888" value="0x02592000" />
+        <register type="NOA" address="0x00009888" value="0x005A8000" />
+        <register type="NOA" address="0x00009888" value="0x025AC000" />
+        <register type="NOA" address="0x00009888" value="0x045A4000" />
+        <register type="NOA" address="0x00009888" value="0x0A1C8000" />
+        <register type="NOA" address="0x00009888" value="0x001D8000" />
+        <register type="NOA" address="0x00009888" value="0x021DC000" />
+        <register type="NOA" address="0x00009888" value="0x041D4000" />
+        <register type="NOA" address="0x00009888" value="0x0A1E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F001A" />
+        <register type="NOA" address="0x00009888" value="0x00394000" />
+        <register type="NOA" address="0x00009888" value="0x02395000" />
+        <register type="NOA" address="0x00009888" value="0x04391000" />
+        <register type="NOA" address="0x00009888" value="0x069A0034" />
+        <register type="NOA" address="0x00009888" value="0x109A0000" />
+        <register type="NOA" address="0x00009888" value="0x06BB4000" />
+        <register type="NOA" address="0x00009888" value="0x0ABE0040" />
+        <register type="NOA" address="0x00009888" value="0x0CBF0984" />
+        <register type="NOA" address="0x00009888" value="0x0EBF0A02" />
+        <register type="NOA" address="0x00009888" value="0x02D94000" />
+        <register type="NOA" address="0x00009888" value="0x0CDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDAC000" />
+        <register type="NOA" address="0x00009888" value="0x0C9C0400" />
+        <register type="NOA" address="0x00009888" value="0x0C9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0E9DC000" />
+        <register type="NOA" address="0x00009888" value="0x0C9E0400" />
+        <register type="NOA" address="0x00009888" value="0x109F02A8" />
+        <register type="NOA" address="0x00009888" value="0x0E9F0040" />
+        <register type="NOA" address="0x00009888" value="0x0CB95000" />
+        <register type="NOA" address="0x00009888" value="0x0EB95000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258B8009" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x198C4000" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185800A" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x1B830154" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x45800800" />
+        <register type="NOA" address="0x00009888" value="0x47800842" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F801084" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800044" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="2c0e45e1-7e2c-4a14-ae00-0b7ec868b8aa"
+       chipset="CHV"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 Rasterizer Input Available"
+             description="The percentage of time in which slice1 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer1_input_available"
+             units="percent"
+             symbol_name="Rasterizer1InputAvailable"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied))"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 Pixel Values Ready"
+             description="The percentage of time in which slice1 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values1_ready"
+             units="percent"
+             symbol_name="PixelValues1Ready"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice1 PS Output Available"
+             description="The percentage of time in which slice1 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output1_available"
+             units="percent"
+             symbol_name="PSOutput1Available"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice1 Rasterizer Output Ready"
+             description="The percentage of time in which slice1 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer1_output_ready"
+             units="percent"
+             symbol_name="Rasterizer1OutputReady"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice1 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice1 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data1_ready"
+             units="percent"
+             symbol_name="PixelData1Ready"
+             availability="$SliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x143B000E" />
+        <register type="NOA" address="0x00009888" value="0x043C55C0" />
+        <register type="NOA" address="0x00009888" value="0x0A1E0280" />
+        <register type="NOA" address="0x00009888" value="0x0C1E0408" />
+        <register type="NOA" address="0x00009888" value="0x10390000" />
+        <register type="NOA" address="0x00009888" value="0x12397A1F" />
+        <register type="NOA" address="0x00009888" value="0x14BB000E" />
+        <register type="NOA" address="0x00009888" value="0x04BC5000" />
+        <register type="NOA" address="0x00009888" value="0x0A9E0296" />
+        <register type="NOA" address="0x00009888" value="0x0C9E0008" />
+        <register type="NOA" address="0x00009888" value="0x10B90000" />
+        <register type="NOA" address="0x00009888" value="0x12B97A1F" />
+        <register type="NOA" address="0x00009888" value="0x063B0042" />
+        <register type="NOA" address="0x00009888" value="0x103B0000" />
+        <register type="NOA" address="0x00009888" value="0x083C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A3E0040" />
+        <register type="NOA" address="0x00009888" value="0x043F8000" />
+        <register type="NOA" address="0x00009888" value="0x02594000" />
+        <register type="NOA" address="0x00009888" value="0x045A8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1C0400" />
+        <register type="NOA" address="0x00009888" value="0x041D8000" />
+        <register type="NOA" address="0x00009888" value="0x081E02C0" />
+        <register type="NOA" address="0x00009888" value="0x0E1E0000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA800" />
+        <register type="NOA" address="0x00009888" value="0x0E1F0260" />
+        <register type="NOA" address="0x00009888" value="0x101F0014" />
+        <register type="NOA" address="0x00009888" value="0x003905E0" />
+        <register type="NOA" address="0x00009888" value="0x06390BC0" />
+        <register type="NOA" address="0x00009888" value="0x02390018" />
+        <register type="NOA" address="0x00009888" value="0x04394000" />
+        <register type="NOA" address="0x00009888" value="0x04BB0042" />
+        <register type="NOA" address="0x00009888" value="0x10BB0000" />
+        <register type="NOA" address="0x00009888" value="0x02BC05C0" />
+        <register type="NOA" address="0x00009888" value="0x08BC0000" />
+        <register type="NOA" address="0x00009888" value="0x0ABE0004" />
+        <register type="NOA" address="0x00009888" value="0x02BF8000" />
+        <register type="NOA" address="0x00009888" value="0x02D91000" />
+        <register type="NOA" address="0x00009888" value="0x02DA8000" />
+        <register type="NOA" address="0x00009888" value="0x089C8000" />
+        <register type="NOA" address="0x00009888" value="0x029D8000" />
+        <register type="NOA" address="0x00009888" value="0x089E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E9E0000" />
+        <register type="NOA" address="0x00009888" value="0x0E9FA806" />
+        <register type="NOA" address="0x00009888" value="0x109F0142" />
+        <register type="NOA" address="0x00009888" value="0x08B90617" />
+        <register type="NOA" address="0x00009888" value="0x0AB90BE0" />
+        <register type="NOA" address="0x00009888" value="0x02B94000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000C" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x018A8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A2800" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x238B52A0" />
+        <register type="NOA" address="0x00009888" value="0x258B6A95" />
+        <register type="NOA" address="0x00009888" value="0x278B0029" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C1500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0014" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x038D8000" />
+        <register type="NOA" address="0x00009888" value="0x058D2000" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA80" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x01834000" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0184C000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1180C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x4D800444" />
+        <register type="NOA" address="0x00009888" value="0x3D800000" />
+        <register type="NOA" address="0x00009888" value="0x4F804000" />
+        <register type="NOA" address="0x00009888" value="0x43801080" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800084" />
+        <register type="NOA" address="0x00009888" value="0x53800044" />
+        <register type="NOA" address="0x00009888" value="0x47801080" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800840" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00006000" />
+        <register type="OA" address="0x00002774" value="0x0000F3FF" />
+        <register type="OA" address="0x00002778" value="0x00001800" />
+        <register type="OA" address="0x0000277C" value="0x0000FCFF" />
+        <register type="OA" address="0x00002780" value="0x00000600" />
+        <register type="OA" address="0x00002784" value="0x0000FF3F" />
+        <register type="OA" address="0x00002788" value="0x00000180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000060" />
+        <register type="OA" address="0x00002794" value="0x0000FFF3" />
+        <register type="OA" address="0x00002798" value="0x00000018" />
+        <register type="OA" address="0x0000279C" value="0x0000FFFC" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler_1"
+       hw_config_guid="71148d78-baf5-474f-878a-e23158d0265d"
+       chipset="CHV"
+       symbol_name="Sampler_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Slice1 Subslice0 Input Available"
+             description="The percentage of time in which slice1 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler10_input_available"
+             units="percent"
+             symbol_name="Sampler10InputAvailable"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Slice1 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice1 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler12_output_ready"
+             units="percent"
+             symbol_name="Sampler12OutputReady"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice1 Subslice1 Input Available"
+             description="The percentage of time in which slice1 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler11_input_available"
+             units="percent"
+             symbol_name="Sampler11InputAvailable"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice1 Subslice2 Input Available"
+             description="The percentage of time in which slice1 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler12_input_available"
+             units="percent"
+             symbol_name="Sampler12InputAvailable"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice1 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice1 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler10_output_ready"
+             units="percent"
+             symbol_name="Sampler10OutputReady"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Slice1 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice1 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler11_output_ready"
+             units="percent"
+             symbol_name="Sampler11OutputReady"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x18921400" />
+        <register type="NOA" address="0x00009888" value="0x149500AB" />
+        <register type="NOA" address="0x00009888" value="0x18B21400" />
+        <register type="NOA" address="0x00009888" value="0x14B500AB" />
+        <register type="NOA" address="0x00009888" value="0x18D21400" />
+        <register type="NOA" address="0x00009888" value="0x14D500AB" />
+        <register type="NOA" address="0x00009888" value="0x0CDC8000" />
+        <register type="NOA" address="0x00009888" value="0x0EDC4000" />
+        <register type="NOA" address="0x00009888" value="0x02DCC000" />
+        <register type="NOA" address="0x00009888" value="0x04DCC000" />
+        <register type="NOA" address="0x00009888" value="0x1ABD00A0" />
+        <register type="NOA" address="0x00009888" value="0x0ABD8000" />
+        <register type="NOA" address="0x00009888" value="0x0CD88000" />
+        <register type="NOA" address="0x00009888" value="0x0ED84000" />
+        <register type="NOA" address="0x00009888" value="0x04D88000" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0050" />
+        <register type="NOA" address="0x00009888" value="0x04DB8000" />
+        <register type="NOA" address="0x00009888" value="0x06DB8000" />
+        <register type="NOA" address="0x00009888" value="0x08DB8000" />
+        <register type="NOA" address="0x00009888" value="0x0ADB4000" />
+        <register type="NOA" address="0x00009888" value="0x109F02A0" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F00AA" />
+        <register type="NOA" address="0x00009888" value="0x18B82500" />
+        <register type="NOA" address="0x00009888" value="0x02B88000" />
+        <register type="NOA" address="0x00009888" value="0x04B84000" />
+        <register type="NOA" address="0x00009888" value="0x06B84000" />
+        <register type="NOA" address="0x00009888" value="0x08B84000" />
+        <register type="NOA" address="0x00009888" value="0x0AB84000" />
+        <register type="NOA" address="0x00009888" value="0x0CB88000" />
+        <register type="NOA" address="0x00009888" value="0x0CB98000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x00B98000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x04B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x1ABA0200" />
+        <register type="NOA" address="0x00009888" value="0x02BA8000" />
+        <register type="NOA" address="0x00009888" value="0x0CBA8000" />
+        <register type="NOA" address="0x00009888" value="0x04908000" />
+        <register type="NOA" address="0x00009888" value="0x04918000" />
+        <register type="NOA" address="0x00009888" value="0x04927300" />
+        <register type="NOA" address="0x00009888" value="0x10920000" />
+        <register type="NOA" address="0x00009888" value="0x1893000A" />
+        <register type="NOA" address="0x00009888" value="0x0A934000" />
+        <register type="NOA" address="0x00009888" value="0x0A946000" />
+        <register type="NOA" address="0x00009888" value="0x0C959000" />
+        <register type="NOA" address="0x00009888" value="0x0E950098" />
+        <register type="NOA" address="0x00009888" value="0x10950000" />
+        <register type="NOA" address="0x00009888" value="0x04B04000" />
+        <register type="NOA" address="0x00009888" value="0x04B14000" />
+        <register type="NOA" address="0x00009888" value="0x04B20073" />
+        <register type="NOA" address="0x00009888" value="0x10B20000" />
+        <register type="NOA" address="0x00009888" value="0x04B38000" />
+        <register type="NOA" address="0x00009888" value="0x06B38000" />
+        <register type="NOA" address="0x00009888" value="0x08B34000" />
+        <register type="NOA" address="0x00009888" value="0x04B4C000" />
+        <register type="NOA" address="0x00009888" value="0x02B59890" />
+        <register type="NOA" address="0x00009888" value="0x10B50000" />
+        <register type="NOA" address="0x00009888" value="0x06D04000" />
+        <register type="NOA" address="0x00009888" value="0x06D14000" />
+        <register type="NOA" address="0x00009888" value="0x06D20073" />
+        <register type="NOA" address="0x00009888" value="0x10D20000" />
+        <register type="NOA" address="0x00009888" value="0x18D30020" />
+        <register type="NOA" address="0x00009888" value="0x02D38000" />
+        <register type="NOA" address="0x00009888" value="0x0CD34000" />
+        <register type="NOA" address="0x00009888" value="0x0AD48000" />
+        <register type="NOA" address="0x00009888" value="0x04D42000" />
+        <register type="NOA" address="0x00009888" value="0x0ED59000" />
+        <register type="NOA" address="0x00009888" value="0x00D59800" />
+        <register type="NOA" address="0x00009888" value="0x10D50000" />
+        <register type="NOA" address="0x00009888" value="0x0F88000E" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B5500" />
+        <register type="NOA" address="0x00009888" value="0x258B000A" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8D8000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x2185000A" />
+        <register type="NOA" address="0x00009888" value="0x1B830150" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D848000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D808000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47801021" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800C64" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800C02" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler_2"
+       hw_config_guid="b996a2b7-c59c-492d-877a-8cd54fd6df84"
+       chipset="CHV"
+       symbol_name="Sampler_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x18121400" />
+        <register type="NOA" address="0x00009888" value="0x141500AB" />
+        <register type="NOA" address="0x00009888" value="0x18321400" />
+        <register type="NOA" address="0x00009888" value="0x143500AB" />
+        <register type="NOA" address="0x00009888" value="0x18521400" />
+        <register type="NOA" address="0x00009888" value="0x145500AB" />
+        <register type="NOA" address="0x00009888" value="0x0C5C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5C4000" />
+        <register type="NOA" address="0x00009888" value="0x025CC000" />
+        <register type="NOA" address="0x00009888" value="0x045CC000" />
+        <register type="NOA" address="0x00009888" value="0x1A3D00A0" />
+        <register type="NOA" address="0x00009888" value="0x0A3D8000" />
+        <register type="NOA" address="0x00009888" value="0x0C588000" />
+        <register type="NOA" address="0x00009888" value="0x0E584000" />
+        <register type="NOA" address="0x00009888" value="0x04588000" />
+        <register type="NOA" address="0x00009888" value="0x1A5B0050" />
+        <register type="NOA" address="0x00009888" value="0x045B8000" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B8000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x101F02A0" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00AA" />
+        <register type="NOA" address="0x00009888" value="0x18382500" />
+        <register type="NOA" address="0x00009888" value="0x02388000" />
+        <register type="NOA" address="0x00009888" value="0x04384000" />
+        <register type="NOA" address="0x00009888" value="0x06384000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0C388000" />
+        <register type="NOA" address="0x00009888" value="0x0C398000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x00398000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x1A3A0200" />
+        <register type="NOA" address="0x00009888" value="0x023A8000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A8000" />
+        <register type="NOA" address="0x00009888" value="0x04108000" />
+        <register type="NOA" address="0x00009888" value="0x04118000" />
+        <register type="NOA" address="0x00009888" value="0x04127300" />
+        <register type="NOA" address="0x00009888" value="0x10120000" />
+        <register type="NOA" address="0x00009888" value="0x1813000A" />
+        <register type="NOA" address="0x00009888" value="0x0A134000" />
+        <register type="NOA" address="0x00009888" value="0x0A146000" />
+        <register type="NOA" address="0x00009888" value="0x0C159000" />
+        <register type="NOA" address="0x00009888" value="0x0E150098" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x04304000" />
+        <register type="NOA" address="0x00009888" value="0x04314000" />
+        <register type="NOA" address="0x00009888" value="0x04320073" />
+        <register type="NOA" address="0x00009888" value="0x10320000" />
+        <register type="NOA" address="0x00009888" value="0x04338000" />
+        <register type="NOA" address="0x00009888" value="0x06338000" />
+        <register type="NOA" address="0x00009888" value="0x08334000" />
+        <register type="NOA" address="0x00009888" value="0x0434C000" />
+        <register type="NOA" address="0x00009888" value="0x02359890" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x06504000" />
+        <register type="NOA" address="0x00009888" value="0x06514000" />
+        <register type="NOA" address="0x00009888" value="0x06520073" />
+        <register type="NOA" address="0x00009888" value="0x10520000" />
+        <register type="NOA" address="0x00009888" value="0x18530020" />
+        <register type="NOA" address="0x00009888" value="0x02538000" />
+        <register type="NOA" address="0x00009888" value="0x0C534000" />
+        <register type="NOA" address="0x00009888" value="0x0A548000" />
+        <register type="NOA" address="0x00009888" value="0x04542000" />
+        <register type="NOA" address="0x00009888" value="0x0E559000" />
+        <register type="NOA" address="0x00009888" value="0x00559800" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AA000" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x258B0005" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x2185000A" />
+        <register type="NOA" address="0x00009888" value="0x1B830150" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0D848000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x07844000" />
+        <register type="NOA" address="0x00009888" value="0x1D808000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x17804000" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47801021" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800C64" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x41800C02" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="eb2fecba-b431-42e7-8261-fe9429a6e67a"
+       chipset="CHV"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice1 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice1 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread11_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread11ReadyForDispatch"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice1 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice1 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread12_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread12ReadyForDispatch"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice1 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice1 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread10_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread10ReadyForDispatch"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice1 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice1 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread12_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread12ReadyForDispatch"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice1 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice1 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread10_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread10ReadyForDispatch"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice1 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice1 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread11_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread11ReadyForDispatch"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x16154D60" />
+        <register type="NOA" address="0x00009888" value="0x16352E60" />
+        <register type="NOA" address="0x00009888" value="0x16554D60" />
+        <register type="NOA" address="0x00009888" value="0x16950000" />
+        <register type="NOA" address="0x00009888" value="0x16B50000" />
+        <register type="NOA" address="0x00009888" value="0x16D50000" />
+        <register type="NOA" address="0x00009888" value="0x005C8000" />
+        <register type="NOA" address="0x00009888" value="0x045CC000" />
+        <register type="NOA" address="0x00009888" value="0x065C4000" />
+        <register type="NOA" address="0x00009888" value="0x083D8000" />
+        <register type="NOA" address="0x00009888" value="0x0A3D8000" />
+        <register type="NOA" address="0x00009888" value="0x0458C000" />
+        <register type="NOA" address="0x00009888" value="0x025B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E1F00AA" />
+        <register type="NOA" address="0x00009888" value="0x02384000" />
+        <register type="NOA" address="0x00009888" value="0x04388000" />
+        <register type="NOA" address="0x00009888" value="0x06388000" />
+        <register type="NOA" address="0x00009888" value="0x08384000" />
+        <register type="NOA" address="0x00009888" value="0x0A384000" />
+        <register type="NOA" address="0x00009888" value="0x0C384000" />
+        <register type="NOA" address="0x00009888" value="0x00398000" />
+        <register type="NOA" address="0x00009888" value="0x0239A000" />
+        <register type="NOA" address="0x00009888" value="0x0439A000" />
+        <register type="NOA" address="0x00009888" value="0x06392000" />
+        <register type="NOA" address="0x00009888" value="0x043A8000" />
+        <register type="NOA" address="0x00009888" value="0x063A8000" />
+        <register type="NOA" address="0x00009888" value="0x08138000" />
+        <register type="NOA" address="0x00009888" value="0x0A138000" />
+        <register type="NOA" address="0x00009888" value="0x06143000" />
+        <register type="NOA" address="0x00009888" value="0x0415CFC7" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x02338000" />
+        <register type="NOA" address="0x00009888" value="0x0C338000" />
+        <register type="NOA" address="0x00009888" value="0x04342000" />
+        <register type="NOA" address="0x00009888" value="0x06344000" />
+        <register type="NOA" address="0x00009888" value="0x0035C700" />
+        <register type="NOA" address="0x00009888" value="0x063500CF" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x04538000" />
+        <register type="NOA" address="0x00009888" value="0x06538000" />
+        <register type="NOA" address="0x00009888" value="0x0454C000" />
+        <register type="NOA" address="0x00009888" value="0x0255CFC7" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x06DC8000" />
+        <register type="NOA" address="0x00009888" value="0x08DC4000" />
+        <register type="NOA" address="0x00009888" value="0x0CDCC000" />
+        <register type="NOA" address="0x00009888" value="0x0EDCC000" />
+        <register type="NOA" address="0x00009888" value="0x1ABD00A8" />
+        <register type="NOA" address="0x00009888" value="0x0CD8C000" />
+        <register type="NOA" address="0x00009888" value="0x0ED84000" />
+        <register type="NOA" address="0x00009888" value="0x0EDB8000" />
+        <register type="NOA" address="0x00009888" value="0x18DB0800" />
+        <register type="NOA" address="0x00009888" value="0x1ADB0254" />
+        <register type="NOA" address="0x00009888" value="0x0E9FAA00" />
+        <register type="NOA" address="0x00009888" value="0x109F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0EB84000" />
+        <register type="NOA" address="0x00009888" value="0x16B84000" />
+        <register type="NOA" address="0x00009888" value="0x18B8156A" />
+        <register type="NOA" address="0x00009888" value="0x06B98000" />
+        <register type="NOA" address="0x00009888" value="0x08B9A000" />
+        <register type="NOA" address="0x00009888" value="0x0AB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0CB9A000" />
+        <register type="NOA" address="0x00009888" value="0x0EB9A000" />
+        <register type="NOA" address="0x00009888" value="0x18BAA000" />
+        <register type="NOA" address="0x00009888" value="0x1ABA0002" />
+        <register type="NOA" address="0x00009888" value="0x16934000" />
+        <register type="NOA" address="0x00009888" value="0x1893000A" />
+        <register type="NOA" address="0x00009888" value="0x0A947000" />
+        <register type="NOA" address="0x00009888" value="0x0C95C5C1" />
+        <register type="NOA" address="0x00009888" value="0x0E9500C3" />
+        <register type="NOA" address="0x00009888" value="0x10950000" />
+        <register type="NOA" address="0x00009888" value="0x0EB38000" />
+        <register type="NOA" address="0x00009888" value="0x16B30040" />
+        <register type="NOA" address="0x00009888" value="0x18B30020" />
+        <register type="NOA" address="0x00009888" value="0x06B48000" />
+        <register type="NOA" address="0x00009888" value="0x08B41000" />
+        <register type="NOA" address="0x00009888" value="0x0AB48000" />
+        <register type="NOA" address="0x00009888" value="0x06B5C500" />
+        <register type="NOA" address="0x00009888" value="0x08B500C3" />
+        <register type="NOA" address="0x00009888" value="0x0EB5C100" />
+        <register type="NOA" address="0x00009888" value="0x10B50000" />
+        <register type="NOA" address="0x00009888" value="0x16D31500" />
+        <register type="NOA" address="0x00009888" value="0x08D4E000" />
+        <register type="NOA" address="0x00009888" value="0x08D5C100" />
+        <register type="NOA" address="0x00009888" value="0x0AD5C3C5" />
+        <register type="NOA" address="0x00009888" value="0x10D50000" />
+        <register type="NOA" address="0x00009888" value="0x0D88F800" />
+        <register type="NOA" address="0x00009888" value="0x0F88000F" />
+        <register type="NOA" address="0x00009888" value="0x038A8000" />
+        <register type="NOA" address="0x00009888" value="0x058A8000" />
+        <register type="NOA" address="0x00009888" value="0x078A8000" />
+        <register type="NOA" address="0x00009888" value="0x098A8000" />
+        <register type="NOA" address="0x00009888" value="0x0B8A8000" />
+        <register type="NOA" address="0x00009888" value="0x0D8A8000" />
+        <register type="NOA" address="0x00009888" value="0x258BAAA5" />
+        <register type="NOA" address="0x00009888" value="0x278B002A" />
+        <register type="NOA" address="0x00009888" value="0x238B2A80" />
+        <register type="NOA" address="0x00009888" value="0x0F8C4000" />
+        <register type="NOA" address="0x00009888" value="0x178C2000" />
+        <register type="NOA" address="0x00009888" value="0x198C5500" />
+        <register type="NOA" address="0x00009888" value="0x1B8C0015" />
+        <register type="NOA" address="0x00009888" value="0x078D8000" />
+        <register type="NOA" address="0x00009888" value="0x098DA000" />
+        <register type="NOA" address="0x00009888" value="0x0B8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0D8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0F8DA000" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800C42" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45800063" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x47800800" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F8014A4" />
+        <register type="NOA" address="0x00009888" value="0x41801042" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x0000FE7F" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFBF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFF7" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FFF9" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="60749470-a648-4a4b-9f10-dbfe1e36e44d"
+       chipset="CHV"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice1 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice1 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header11_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader11ReadyPort1"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice1 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header11_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader11ReadyPort0"
+             availability="$SubsliceMask 0x10 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice1 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header12_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader12ReadyPort0"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice1 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header10_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader10ReadyPort1"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice1 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header12_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader12ReadyPort1"
+             availability="$SubsliceMask 0x20 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL 2 UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice1 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice1 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header10_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader10ReadyPort0"
+             availability="$SubsliceMask 0x8 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x16150000" />
+        <register type="NOA" address="0x00009888" value="0x16350000" />
+        <register type="NOA" address="0x00009888" value="0x16550000" />
+        <register type="NOA" address="0x00009888" value="0x16952E60" />
+        <register type="NOA" address="0x00009888" value="0x16B54D60" />
+        <register type="NOA" address="0x00009888" value="0x16D52E60" />
+        <register type="NOA" address="0x00009888" value="0x065C8000" />
+        <register type="NOA" address="0x00009888" value="0x085CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A5CC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5C4000" />
+        <register type="NOA" address="0x00009888" value="0x0E3D8000" />
+        <register type="NOA" address="0x00009888" value="0x183DA000" />
+        <register type="NOA" address="0x00009888" value="0x06588000" />
+        <register type="NOA" address="0x00009888" value="0x08588000" />
+        <register type="NOA" address="0x00009888" value="0x0A584000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x185B5800" />
+        <register type="NOA" address="0x00009888" value="0x1A5B000A" />
+        <register type="NOA" address="0x00009888" value="0x0E1FAA00" />
+        <register type="NOA" address="0x00009888" value="0x101F02AA" />
+        <register type="NOA" address="0x00009888" value="0x0E384000" />
+        <register type="NOA" address="0x00009888" value="0x16384000" />
+        <register type="NOA" address="0x00009888" value="0x18382A55" />
+        <register type="NOA" address="0x00009888" value="0x06398000" />
+        <register type="NOA" address="0x00009888" value="0x0839A000" />
+        <register type="NOA" address="0x00009888" value="0x0A39A000" />
+        <register type="NOA" address="0x00009888" value="0x0C39A000" />
+        <register type="NOA" address="0x00009888" value="0x0E39A000" />
+        <register type="NOA" address="0x00009888" value="0x1A3A02A0" />
+        <register type="NOA" address="0x00009888" value="0x0E138000" />
+        <register type="NOA" address="0x00009888" value="0x16130500" />
+        <register type="NOA" address="0x00009888" value="0x06148000" />
+        <register type="NOA" address="0x00009888" value="0x08146000" />
+        <register type="NOA" address="0x00009888" value="0x0615C100" />
+        <register type="NOA" address="0x00009888" value="0x0815C500" />
+        <register type="NOA" address="0x00009888" value="0x0A1500C3" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x16335040" />
+        <register type="NOA" address="0x00009888" value="0x08349000" />
+        <register type="NOA" address="0x00009888" value="0x0A341000" />
+        <register type="NOA" address="0x00009888" value="0x083500C1" />
+        <register type="NOA" address="0x00009888" value="0x0A35C500" />
+        <register type="NOA" address="0x00009888" value="0x0C3500C3" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x1853002A" />
+        <register type="NOA" address="0x00009888" value="0x0A54E000" />
+        <register type="NOA" address="0x00009888" value="0x0C55C500" />
+        <register type="NOA" address="0x00009888" value="0x0E55C1C3" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x00DC8000" />
+        <register type="NOA" address="0x00009888" value="0x02DCC000" />
+        <register type="NOA" address="0x00009888" value="0x04DC4000" />
+        <register type="NOA" address="0x00009888" value="0x04BD8000" />
+        <register type="NOA" address="0x00009888" value="0x06BD8000" />
+        <register type="NOA" address="0x00009888" value="0x02D8C000" />
+        <register type="NOA" address="0x00009888" value="0x02DB8000" />
+        <register type="NOA" address="0x00009888" value="0x04DB4000" />
+        <register type="NOA" address="0x00009888" value="0x06DB4000" />
+        <register type="NOA" address="0x00009888" value="0x08DB8000" />
+        <register type="NOA" address="0x00009888" value="0x0C9FA000" />
+        <register type="NOA" address="0x00009888" value="0x0E9F00AA" />
+        <register type="NOA" address="0x00009888" value="0x02B84000" />
+        <register type="NOA" address="0x00009888" value="0x04B84000" />
+        <register type="NOA" address="0x00009888" value="0x06B84000" />
+        <register type="NOA" address="0x00009888" value="0x08B84000" />
+        <register type="NOA" address="0x00009888" value="0x0AB88000" />
+        <register type="NOA" address="0x00009888" value="0x0CB88000" />
+        <register type="NOA" address="0x00009888" value="0x00B98000" />
+        <register type="NOA" address="0x00009888" value="0x02B9A000" />
+        <register type="NOA" address="0x00009888" value="0x04B9A000" />
+        <register type="NOA" address="0x00009888" value="0x06B92000" />
+        <register type="NOA" address="0x00009888" value="0x0ABA8000" />
+        <register type="NOA" address="0x00009888" value="0x0CBA8000" />
+        <register type="NOA" address="0x00009888" value="0x04938000" />
+        <register type="NOA" address="0x00009888" value="0x06938000" />
+        <register type="NOA" address="0x00009888" value="0x0494C000" />
+        <register type="NOA" address="0x00009888" value="0x0295CFC7" />
+        <register type="NOA" address="0x00009888" value="0x10950000" />
+        <register type="NOA" address="0x00009888" value="0x02B38000" />
+        <register type="NOA" address="0x00009888" value="0x08B38000" />
+        <register type="NOA" address="0x00009888" value="0x04B42000" />
+        <register type="NOA" address="0x00009888" value="0x06B41000" />
+        <register type="NOA" address="0x00009888" value="0x00B5C700" />
+        <register type="NOA" address="0x00009888" value="0x04B500CF" />
+        <register type="NOA" address="0x00009888" value="0x10B50000" />
+        <register type="NOA" address="0x00009888" value="0x0AD38000" />
+        <register type="NOA" address="0x00009888" value="0x0CD38000" />
+        <register type="NOA" address="0x00009888" value="0x06D46000" />
+        <register type="NOA" address="0x00009888" value="0x04D5C700" />
+        <register type="NOA" address="0x00009888" value="0x06D500CF" />
+        <register type="NOA" address="0x00009888" value="0x10D50000" />
+        <register type="NOA" address="0x00009888" value="0x03888000" />
+        <register type="NOA" address="0x00009888" value="0x05888000" />
+        <register type="NOA" address="0x00009888" value="0x07888000" />
+        <register type="NOA" address="0x00009888" value="0x09888000" />
+        <register type="NOA" address="0x00009888" value="0x0B888000" />
+        <register type="NOA" address="0x00009888" value="0x0D880400" />
+        <register type="NOA" address="0x00009888" value="0x0F8A8000" />
+        <register type="NOA" address="0x00009888" value="0x198A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8AAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0002" />
+        <register type="NOA" address="0x00009888" value="0x258B555A" />
+        <register type="NOA" address="0x00009888" value="0x278B0015" />
+        <register type="NOA" address="0x00009888" value="0x238B5500" />
+        <register type="NOA" address="0x00009888" value="0x038C4000" />
+        <register type="NOA" address="0x00009888" value="0x058C4000" />
+        <register type="NOA" address="0x00009888" value="0x078C4000" />
+        <register type="NOA" address="0x00009888" value="0x098C4000" />
+        <register type="NOA" address="0x00009888" value="0x0B8C4000" />
+        <register type="NOA" address="0x00009888" value="0x0D8C4000" />
+        <register type="NOA" address="0x00009888" value="0x018D8000" />
+        <register type="NOA" address="0x00009888" value="0x038DA000" />
+        <register type="NOA" address="0x00009888" value="0x058DA000" />
+        <register type="NOA" address="0x00009888" value="0x078D2000" />
+        <register type="NOA" address="0x00009888" value="0x2185AAAA" />
+        <register type="NOA" address="0x00009888" value="0x2385002A" />
+        <register type="NOA" address="0x00009888" value="0x1F85AA00" />
+        <register type="NOA" address="0x00009888" value="0x0F834000" />
+        <register type="NOA" address="0x00009888" value="0x19835400" />
+        <register type="NOA" address="0x00009888" value="0x1B830155" />
+        <register type="NOA" address="0x00009888" value="0x03834000" />
+        <register type="NOA" address="0x00009888" value="0x05834000" />
+        <register type="NOA" address="0x00009888" value="0x07834000" />
+        <register type="NOA" address="0x00009888" value="0x09834000" />
+        <register type="NOA" address="0x00009888" value="0x0B834000" />
+        <register type="NOA" address="0x00009888" value="0x0D834000" />
+        <register type="NOA" address="0x00009888" value="0x0784C000" />
+        <register type="NOA" address="0x00009888" value="0x0984C000" />
+        <register type="NOA" address="0x00009888" value="0x0B84C000" />
+        <register type="NOA" address="0x00009888" value="0x0D84C000" />
+        <register type="NOA" address="0x00009888" value="0x0F84C000" />
+        <register type="NOA" address="0x00009888" value="0x01848000" />
+        <register type="NOA" address="0x00009888" value="0x0384C000" />
+        <register type="NOA" address="0x00009888" value="0x0584C000" />
+        <register type="NOA" address="0x00009888" value="0x1780C000" />
+        <register type="NOA" address="0x00009888" value="0x1980C000" />
+        <register type="NOA" address="0x00009888" value="0x1B80C000" />
+        <register type="NOA" address="0x00009888" value="0x1D80C000" />
+        <register type="NOA" address="0x00009888" value="0x1F80C000" />
+        <register type="NOA" address="0x00009888" value="0x11808000" />
+        <register type="NOA" address="0x00009888" value="0x1380C000" />
+        <register type="NOA" address="0x00009888" value="0x1580C000" />
+        <register type="NOA" address="0x00009888" value="0x4F800000" />
+        <register type="NOA" address="0x00009888" value="0x43800882" />
+        <register type="NOA" address="0x00009888" value="0x51800000" />
+        <register type="NOA" address="0x00009888" value="0x45801082" />
+        <register type="NOA" address="0x00009888" value="0x53800000" />
+        <register type="NOA" address="0x00009888" value="0x478014A5" />
+        <register type="NOA" address="0x00009888" value="0x21800000" />
+        <register type="NOA" address="0x00009888" value="0x31800000" />
+        <register type="NOA" address="0x00009888" value="0x4D800000" />
+        <register type="NOA" address="0x00009888" value="0x3F800002" />
+        <register type="NOA" address="0x00009888" value="0x41800C62" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x0000FE7F" />
+        <register type="OA" address="0x00002780" value="0x00000000" />
+        <register type="OA" address="0x00002784" value="0x0000FF9F" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000FFE7" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFFB" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FFFD" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen8LP"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="4a534b07-cba3-414d-8d60-874830e883aa"
+       chipset="CHV"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+        <register type="NOA" address="0x00009888" value="0x59800001" />
+        <register type="NOA" address="0x00009888" value="0x338B0000" />
+        <register type="NOA" address="0x00009888" value="0x258B0066" />
+        <register type="NOA" address="0x00009888" value="0x058B0000" />
+        <register type="NOA" address="0x00009888" value="0x038B0000" />
+        <register type="NOA" address="0x00009888" value="0x03844000" />
+        <register type="NOA" address="0x00009888" value="0x47800080" />
+        <register type="NOA" address="0x00009888" value="0x57800000" />
+        <register type="NOA" address="0x001823A4" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x59800000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
diff --git a/src/mesa/drivers/dri/i965/brw_oa_hsw.xml b/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
index 739ea0e..a227640 100644
--- a/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
+++ b/src/mesa/drivers/dri/i965/brw_oa_hsw.xml
@@ -1,5 +1,5 @@
 <?xml version="1.0"?>
-<metrics version="1486658299" merge_md5="">
+<metrics version="1490233727" merge_md5="">
   <set name="Render Metrics Basic Gen7.5"
        mdapi_supported_apis="OGL OCL IO BB"
        underscore_name="render_basic"
@@ -926,7 +926,7 @@
              mdapi_hw_unit_type="gpu"
              mdapi_group="3D Pipe/Output Merger"
              />
-    <register_config id="0">
+    <register_config type="NOA">
         <register type="NOA" address="0x000253A4" value="0x01600000" />
         <register type="NOA" address="0x00025440" value="0x00100000" />
         <register type="NOA" address="0x00025128" value="0x00000000" />
@@ -987,7 +987,7 @@
         <register type="NOA" address="0x0002541C" value="0x00000000" />
         <register type="NOA" address="0x00025428" value="0x00042049" />
     </register_config>
-    <register_config id="1">
+    <register_config type="OA">
         <register type="OA" address="0x00002724" value="0x00800000" /> <!--OASTARTTRIG6 -->
         <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
         <register type="OA" address="0x00002714" value="0x00800000" /> <!--OASTARTTRIG1 -->
@@ -1680,7 +1680,7 @@
              mdapi_group="EU Array/Vertex Shader"
              mdapi_hw_unit_type="gpu"
              />
-    <register_config id="0">
+    <register_config type="NOA">
         <register type="NOA" address="0x000253A4" value="0x00000000" />
         <register type="NOA" address="0x0002681C" value="0x01F00800" />
         <register type="NOA" address="0x00026820" value="0x00001000" />
@@ -1714,7 +1714,7 @@
         <register type="NOA" address="0x0002541C" value="0x00000000" />
         <register type="NOA" address="0x00025428" value="0x00000C03" />
     </register_config>
-    <register_config id="1">
+    <register_config type="OA">
         <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
         <register type="OA" address="0x00002714" value="0x00800000" /> <!--OASTARTTRIG1 -->
         <register type="OA" address="0x00002718" value="0xAAAAAAAA" /> <!--OASTARTTRIG1 -->
@@ -2002,7 +2002,7 @@
              mdapi_hw_unit_type="subslice"
              mdapi_group="L3/Data Port"
              />
-    <register_config id="0">
+    <register_config type="NOA">
         <register type="NOA" address="0x0002681C" value="0x3EB00800" />
         <register type="NOA" address="0x00026820" value="0x00900000" />
         <register type="NOA" address="0x00025384" value="0x02AAAAAA" />
@@ -2019,7 +2019,7 @@
         <register type="NOA" address="0x0002541C" value="0x00000000" />
         <register type="NOA" address="0x00025428" value="0x00000000" />
     </register_config>
-    <register_config id="1">
+    <register_config type="OA">
         <register type="OA" address="0x00002724" value="0xF0800000" /> <!--OASTARTTRIG6 -->
         <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
         <register type="OA" address="0x00002714" value="0xF0800000" /> <!--OASTARTTRIG1 -->
@@ -2774,7 +2774,7 @@
              mdapi_group="EU Array/Vertex Shader"
              mdapi_hw_unit_type="gpu"
              />
-    <register_config id="0">
+    <register_config type="NOA">
         <register type="NOA" address="0x000253A4" value="0x34300000" />
         <register type="NOA" address="0x00025440" value="0x2D800000" />
         <register type="NOA" address="0x00025444" value="0x00000008" />
@@ -2794,7 +2794,7 @@
         <register type="NOA" address="0x0002541C" value="0x00000000" />
         <register type="NOA" address="0x00025428" value="0x00000000" />
     </register_config>
-    <register_config id="1">
+    <register_config type="OA">
         <register type="OA" address="0x00002724" value="0xF0800000" /> <!--OASTARTTRIG6 -->
         <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
         <register type="OA" address="0x00002714" value="0xF0800000" /> <!--OASTARTTRIG1 -->
@@ -3544,7 +3544,7 @@
              mdapi_group="EU Array/Vertex Shader"
              mdapi_hw_unit_type="gpu"
              />
-    <register_config id="0">
+    <register_config type="NOA">
         <register type="NOA" address="0x000253A4" value="0x34300000" />
         <register type="NOA" address="0x00025440" value="0x01500000" />
         <register type="NOA" address="0x00025444" value="0x00000120" />
@@ -3564,7 +3564,7 @@
         <register type="NOA" address="0x0002541C" value="0x00000000" />
         <register type="NOA" address="0x00025428" value="0x00000000" />
     </register_config>
-    <register_config id="1">
+    <register_config type="OA">
         <register type="OA" address="0x00002724" value="0xF0800000" /> <!--OASTARTTRIG6 -->
         <register type="OA" address="0x00002720" value="0x00000000" /> <!--OASTARTTRIG1 -->
         <register type="OA" address="0x00002714" value="0xF0800000" /> <!--OASTARTTRIG1 -->
@@ -4345,7 +4345,7 @@
              mdapi_hw_unit_type="gpu"
              mdapi_group="3D Pipe/Output Merger"
              />
-    <register_config id="0">
+    <register_config type="NOA">
         <register type="NOA" address="0x0002EB9C" value="0x01906400" />
         <register type="NOA" address="0x0002FB9C" value="0x01906400" />
         <register type="NOA" address="0x000253A4" value="0x00000000" />
@@ -4387,7 +4387,7 @@
         <register type="NOA" address="0x0002541C" value="0x00000000" />
         <register type="NOA" address="0x00025428" value="0x0004A54A" />
     </register_config>
-    <register_config id="1">
+    <register_config type="OA">
         <register type="OA" address="0x00002740" value="0x00000000" /> <!--OAREPORTTRIG1 -->
         <register type="OA" address="0x00002744" value="0x00800000" /> <!--OAREPORTTRIG2 -->
         <register type="OA" address="0x00002710" value="0x00000000" /> <!--OASTARTTRIG1 -->
diff --git a/src/mesa/drivers/dri/i965/brw_oa_sklgt2.xml b/src/mesa/drivers/dri/i965/brw_oa_sklgt2.xml
new file mode 100644
index 0000000..eb28450
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_sklgt2.xml
@@ -0,0 +1,10925 @@
+<?xml version="1.0"?>
+<metrics version="1491577975" merge_md5="">
+  <set name="Render Metrics Basic Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="f519e481-24d2-4d42-87c9-3fdd12c00202"
+       chipset="SKLGT2"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Cache Misses"
+             description="The total number of sampler cache misses in all LODs in all sampler units."
+             data_type="uint64"
+             equation="B 4 READ 8 UMUL"
+             underscore_name="sampler_l1_misses"
+             units="messages"
+             symbol_name="SamplerL1Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 1 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$SamplerL1Misses 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="$Sampler0Busy $Sampler1Busy FMAX"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="L3 Lookup Accesses w/o IC"
+             description="The total number of L3 cache lookup accesses w/o IC."
+             data_type="uint64"
+             equation="$SamplerL1Misses $ShaderMemoryAccesses UADD"
+             underscore_name="l3_lookups"
+             units="messages"
+             symbol_name="L3Lookups"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x02 UGTE"
+                     priority="1"
+                     >
+        <register type="NOA" address="0x00009888" value="0x166C01E0" />
+        <register type="NOA" address="0x00009888" value="0x12170280" />
+        <register type="NOA" address="0x00009888" value="0x12370280" />
+        <register type="NOA" address="0x00009888" value="0x11930317" />
+        <register type="NOA" address="0x00009888" value="0x159303DF" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0080" />
+        <register type="NOA" address="0x00009888" value="0x0A6C0053" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A1B4000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0001" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x042F1000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8400" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F6600" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C2200" />
+        <register type="NOA" address="0x00009888" value="0x062D8000" />
+        <register type="NOA" address="0x00009888" value="0x082D8000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x08133000" />
+        <register type="NOA" address="0x00009888" value="0x00170020" />
+        <register type="NOA" address="0x00009888" value="0x08170021" />
+        <register type="NOA" address="0x00009888" value="0x10170000" />
+        <register type="NOA" address="0x00009888" value="0x0633C000" />
+        <register type="NOA" address="0x00009888" value="0x0833C000" />
+        <register type="NOA" address="0x00009888" value="0x06370800" />
+        <register type="NOA" address="0x00009888" value="0x08370840" />
+        <register type="NOA" address="0x00009888" value="0x10370000" />
+        <register type="NOA" address="0x00009888" value="0x0D933031" />
+        <register type="NOA" address="0x00009888" value="0x0F933E3F" />
+        <register type="NOA" address="0x00009888" value="0x01933D00" />
+        <register type="NOA" address="0x00009888" value="0x0393073C" />
+        <register type="NOA" address="0x00009888" value="0x0593000E" />
+        <register type="NOA" address="0x00009888" value="0x1D930000" />
+        <register type="NOA" address="0x00009888" value="0x19930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x2F908000" />
+        <register type="NOA" address="0x00009888" value="0x31908000" />
+        <register type="NOA" address="0x00009888" value="0x15908000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190001F" />
+        <register type="NOA" address="0x00009888" value="0x51904400" />
+        <register type="NOA" address="0x00009888" value="0x41900020" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C21" />
+        <register type="NOA" address="0x00009888" value="0x47900061" />
+        <register type="NOA" address="0x00009888" value="0x57904440" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900004" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x53904444" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="fe47b29d-ae51-423e-bff4-27d965a95b60"
+       chipset="SKLGT2"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ C 0 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 3 READ B 4 READ B 5 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 1 READ C 2 READ C 3 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ B 2 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA"
+                     availability="$SliceMask 0x01 AND $SkuRevisionId 0x02 ULT &amp;&amp;"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x124F1C00" />
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901403" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E8200" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x004F0DB2" />
+        <register type="NOA" address="0x00009888" value="0x064F0900" />
+        <register type="NOA" address="0x00009888" value="0x084F1880" />
+        <register type="NOA" address="0x00009888" value="0x0A4F0011" />
+        <register type="NOA" address="0x00009888" value="0x0C4F0E3C" />
+        <register type="NOA" address="0x00009888" value="0x0E4F1D80" />
+        <register type="NOA" address="0x00009888" value="0x086C0002" />
+        <register type="NOA" address="0x00009888" value="0x0A6C0100" />
+        <register type="NOA" address="0x00009888" value="0x0E6C000C" />
+        <register type="NOA" address="0x00009888" value="0x026C000B" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x081B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A1B8000" />
+        <register type="NOA" address="0x00009888" value="0x0E1B4000" />
+        <register type="NOA" address="0x00009888" value="0x021B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C4000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0012" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x005BC000" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B8000" />
+        <register type="NOA" address="0x00009888" value="0x0A5B4000" />
+        <register type="NOA" address="0x00009888" value="0x0C5BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B8000" />
+        <register type="NOA" address="0x00009888" value="0x105C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A5CA000" />
+        <register type="NOA" address="0x00009888" value="0x1C5C002D" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0800" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0082" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002CC000" />
+        <register type="NOA" address="0x00009888" value="0x0E2CC000" />
+        <register type="NOA" address="0x00009888" value="0x162CBE00" />
+        <register type="NOA" address="0x00009888" value="0x182C00EF" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x19900157" />
+        <register type="NOA" address="0x00009888" value="0x1B900167" />
+        <register type="NOA" address="0x00009888" value="0x1D900105" />
+        <register type="NOA" address="0x00009888" value="0x1F900103" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00000D28" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x11900FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900840" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900842" />
+        <register type="NOA" address="0x00009888" value="0x47900840" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900840" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900040" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900840" />
+        <register type="NOA" address="0x00009888" value="0x53901111" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SliceMask 0x01 AND $SkuRevisionId 0x02 UGTE &amp;&amp;"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x124F1C00" />
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901403" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0820" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x064F0900" />
+        <register type="NOA" address="0x00009888" value="0x084F0032" />
+        <register type="NOA" address="0x00009888" value="0x0A4F1810" />
+        <register type="NOA" address="0x00009888" value="0x0C4F0E00" />
+        <register type="NOA" address="0x00009888" value="0x0E4F003C" />
+        <register type="NOA" address="0x00009888" value="0x004F0D80" />
+        <register type="NOA" address="0x00009888" value="0x024F003B" />
+        <register type="NOA" address="0x00009888" value="0x006C0002" />
+        <register type="NOA" address="0x00009888" value="0x086C0000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C000C" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B00" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x081B8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E1B8000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C8000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0024" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5BC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C6000" />
+        <register type="NOA" address="0x00009888" value="0x1C5C001B" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0208" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2CC000" />
+        <register type="NOA" address="0x00009888" value="0x162CFB00" />
+        <register type="NOA" address="0x00009888" value="0x182C00BE" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x19900157" />
+        <register type="NOA" address="0x00009888" value="0x1B900167" />
+        <register type="NOA" address="0x00009888" value="0x1D900105" />
+        <register type="NOA" address="0x00009888" value="0x1F900103" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x11900FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900842" />
+        <register type="NOA" address="0x00009888" value="0x47900802" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900802" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900002" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53901111" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="e0ad5ae0-84ba-4f29-a723-1906c12cb774"
+       chipset="SKLGT2"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x02 ULT"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x0C0E001F" />
+        <register type="NOA" address="0x00009888" value="0x0A0F0000" />
+        <register type="NOA" address="0x00009888" value="0x10116800" />
+        <register type="NOA" address="0x00009888" value="0x178A03E0" />
+        <register type="NOA" address="0x00009888" value="0x11824C00" />
+        <register type="NOA" address="0x00009888" value="0x11830020" />
+        <register type="NOA" address="0x00009888" value="0x13840020" />
+        <register type="NOA" address="0x00009888" value="0x11850019" />
+        <register type="NOA" address="0x00009888" value="0x11860007" />
+        <register type="NOA" address="0x00009888" value="0x01870C40" />
+        <register type="NOA" address="0x00009888" value="0x17880000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x040D4000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020E5400" />
+        <register type="NOA" address="0x00009888" value="0x000E0000" />
+        <register type="NOA" address="0x00009888" value="0x080F0040" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x100F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0040" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06110012" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x01898000" />
+        <register type="NOA" address="0x00009888" value="0x0D890100" />
+        <register type="NOA" address="0x00009888" value="0x03898000" />
+        <register type="NOA" address="0x00009888" value="0x09808000" />
+        <register type="NOA" address="0x00009888" value="0x0B808000" />
+        <register type="NOA" address="0x00009888" value="0x0380C000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A0075" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0000" />
+        <register type="NOA" address="0x00009888" value="0x118A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A4000" />
+        <register type="NOA" address="0x00009888" value="0x138A8000" />
+        <register type="NOA" address="0x00009888" value="0x1D81A000" />
+        <register type="NOA" address="0x00009888" value="0x15818000" />
+        <register type="NOA" address="0x00009888" value="0x17818000" />
+        <register type="NOA" address="0x00009888" value="0x0B820030" />
+        <register type="NOA" address="0x00009888" value="0x07828000" />
+        <register type="NOA" address="0x00009888" value="0x0D824000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x05824000" />
+        <register type="NOA" address="0x00009888" value="0x0D830003" />
+        <register type="NOA" address="0x00009888" value="0x0583000C" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x03838000" />
+        <register type="NOA" address="0x00009888" value="0x07838000" />
+        <register type="NOA" address="0x00009888" value="0x0B840980" />
+        <register type="NOA" address="0x00009888" value="0x03844D80" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x09850080" />
+        <register type="NOA" address="0x00009888" value="0x03850003" />
+        <register type="NOA" address="0x00009888" value="0x01850000" />
+        <register type="NOA" address="0x00009888" value="0x07860000" />
+        <register type="NOA" address="0x00009888" value="0x0F860400" />
+        <register type="NOA" address="0x00009888" value="0x09870032" />
+        <register type="NOA" address="0x00009888" value="0x01888052" />
+        <register type="NOA" address="0x00009888" value="0x11880000" />
+        <register type="NOA" address="0x00009888" value="0x09884000" />
+        <register type="NOA" address="0x00009888" value="0x15968000" />
+        <register type="NOA" address="0x00009888" value="0x17968000" />
+        <register type="NOA" address="0x00009888" value="0x0F96C000" />
+        <register type="NOA" address="0x00009888" value="0x1F950011" />
+        <register type="NOA" address="0x00009888" value="0x1D950014" />
+        <register type="NOA" address="0x00009888" value="0x0592C000" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D924000" />
+        <register type="NOA" address="0x00009888" value="0x0F924000" />
+        <register type="NOA" address="0x00009888" value="0x11928000" />
+        <register type="NOA" address="0x00009888" value="0x1392C000" />
+        <register type="NOA" address="0x00009888" value="0x09924000" />
+        <register type="NOA" address="0x00009888" value="0x01985000" />
+        <register type="NOA" address="0x00009888" value="0x07988000" />
+        <register type="NOA" address="0x00009888" value="0x09981000" />
+        <register type="NOA" address="0x00009888" value="0x0B982000" />
+        <register type="NOA" address="0x00009888" value="0x0D982000" />
+        <register type="NOA" address="0x00009888" value="0x0F989000" />
+        <register type="NOA" address="0x00009888" value="0x05982000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x0B978000" />
+        <register type="NOA" address="0x00009888" value="0x0F974000" />
+        <register type="NOA" address="0x00009888" value="0x11974000" />
+        <register type="NOA" address="0x00009888" value="0x13978000" />
+        <register type="NOA" address="0x00009888" value="0x09974000" />
+        <register type="NOA" address="0x00000D28" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x1190C080" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x419010A0" />
+        <register type="NOA" address="0x00009888" value="0x55904000" />
+        <register type="NOA" address="0x00009888" value="0x45901000" />
+        <register type="NOA" address="0x00009888" value="0x47900084" />
+        <register type="NOA" address="0x00009888" value="0x57904400" />
+        <register type="NOA" address="0x00009888" value="0x499000A5" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900081" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x439014A4" />
+        <register type="NOA" address="0x00009888" value="0x53900400" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x02 UGTE"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x0C0E001F" />
+        <register type="NOA" address="0x00009888" value="0x0A0F0000" />
+        <register type="NOA" address="0x00009888" value="0x10116800" />
+        <register type="NOA" address="0x00009888" value="0x178A03E0" />
+        <register type="NOA" address="0x00009888" value="0x11824C00" />
+        <register type="NOA" address="0x00009888" value="0x11830020" />
+        <register type="NOA" address="0x00009888" value="0x13840020" />
+        <register type="NOA" address="0x00009888" value="0x11850019" />
+        <register type="NOA" address="0x00009888" value="0x11860007" />
+        <register type="NOA" address="0x00009888" value="0x01870C40" />
+        <register type="NOA" address="0x00009888" value="0x17880000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x040D4000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020E5400" />
+        <register type="NOA" address="0x00009888" value="0x000E0000" />
+        <register type="NOA" address="0x00009888" value="0x080F0040" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x100F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0040" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06110012" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x01898000" />
+        <register type="NOA" address="0x00009888" value="0x0D890100" />
+        <register type="NOA" address="0x00009888" value="0x03898000" />
+        <register type="NOA" address="0x00009888" value="0x09808000" />
+        <register type="NOA" address="0x00009888" value="0x0B808000" />
+        <register type="NOA" address="0x00009888" value="0x0380C000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A0075" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0000" />
+        <register type="NOA" address="0x00009888" value="0x118A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A4000" />
+        <register type="NOA" address="0x00009888" value="0x138A8000" />
+        <register type="NOA" address="0x00009888" value="0x1D81A000" />
+        <register type="NOA" address="0x00009888" value="0x15818000" />
+        <register type="NOA" address="0x00009888" value="0x17818000" />
+        <register type="NOA" address="0x00009888" value="0x0B820030" />
+        <register type="NOA" address="0x00009888" value="0x07828000" />
+        <register type="NOA" address="0x00009888" value="0x0D824000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x05824000" />
+        <register type="NOA" address="0x00009888" value="0x0D830003" />
+        <register type="NOA" address="0x00009888" value="0x0583000C" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x03838000" />
+        <register type="NOA" address="0x00009888" value="0x07838000" />
+        <register type="NOA" address="0x00009888" value="0x0B840980" />
+        <register type="NOA" address="0x00009888" value="0x03844D80" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x09850080" />
+        <register type="NOA" address="0x00009888" value="0x03850003" />
+        <register type="NOA" address="0x00009888" value="0x01850000" />
+        <register type="NOA" address="0x00009888" value="0x07860000" />
+        <register type="NOA" address="0x00009888" value="0x0F860400" />
+        <register type="NOA" address="0x00009888" value="0x09870032" />
+        <register type="NOA" address="0x00009888" value="0x01888052" />
+        <register type="NOA" address="0x00009888" value="0x11880000" />
+        <register type="NOA" address="0x00009888" value="0x09884000" />
+        <register type="NOA" address="0x00009888" value="0x1B931001" />
+        <register type="NOA" address="0x00009888" value="0x1D930001" />
+        <register type="NOA" address="0x00009888" value="0x19934000" />
+        <register type="NOA" address="0x00009888" value="0x1B958000" />
+        <register type="NOA" address="0x00009888" value="0x1D950094" />
+        <register type="NOA" address="0x00009888" value="0x19958000" />
+        <register type="NOA" address="0x00009888" value="0x05E5A000" />
+        <register type="NOA" address="0x00009888" value="0x01E5C000" />
+        <register type="NOA" address="0x00009888" value="0x0592C000" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D924000" />
+        <register type="NOA" address="0x00009888" value="0x0F924000" />
+        <register type="NOA" address="0x00009888" value="0x11928000" />
+        <register type="NOA" address="0x00009888" value="0x1392C000" />
+        <register type="NOA" address="0x00009888" value="0x09924000" />
+        <register type="NOA" address="0x00009888" value="0x01985000" />
+        <register type="NOA" address="0x00009888" value="0x07988000" />
+        <register type="NOA" address="0x00009888" value="0x09981000" />
+        <register type="NOA" address="0x00009888" value="0x0B982000" />
+        <register type="NOA" address="0x00009888" value="0x0D982000" />
+        <register type="NOA" address="0x00009888" value="0x0F989000" />
+        <register type="NOA" address="0x00009888" value="0x05982000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1190C080" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x419010A0" />
+        <register type="NOA" address="0x00009888" value="0x55904000" />
+        <register type="NOA" address="0x00009888" value="0x45901000" />
+        <register type="NOA" address="0x00009888" value="0x47900084" />
+        <register type="NOA" address="0x00009888" value="0x57904400" />
+        <register type="NOA" address="0x00009888" value="0x499000A5" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900081" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x439014A4" />
+        <register type="NOA" address="0x00009888" value="0x53900400" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="9bc436dd-6130-4add-affc-283eb6eaa864"
+       chipset="SKLGT2"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank0Reads"
+             description="The total number of GTI memory reads from L3 Bank 0 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_reads"
+             units="messages"
+             symbol_name="GtiL3Bank0Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all accesses from GTI to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiL3Bank3Reads"
+             description="The total number of GTI memory reads from L3 Bank 3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_reads"
+             units="messages"
+             symbol_name="GtiL3Bank3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiRsMemoryReads"
+             description="The total number of GTI memory reads from Resource Streamer."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_rs_memory_reads"
+             units="messages"
+             symbol_name="GtiRsMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiHizMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache (Hi-Depth Cache misses)."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_reads"
+             units="messages"
+             symbol_name="GtiHizMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="GtiRccMemoryReads"
+             description="The total number of GTI memory reads from Render Color Cache (Render Color Cache misses)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_reads"
+             units="messages"
+             symbol_name="GtiRccMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank1Reads"
+             description="The total number of GTI memory reads from L3 Bank 1 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_reads"
+             units="messages"
+             symbol_name="GtiL3Bank1Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiCmdStreamerMemoryReads"
+             description="The total number of GTI memory reads from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_reads"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GtiL3Bank2Reads"
+             description="The total number of GTI memory reads from L3 Bank 2 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_reads"
+             units="messages"
+             symbol_name="GtiL3Bank2Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiMemoryReads"
+             description="The total number of GTI memory reads."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache (Render Depth Cache misses)."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="GtiMscMemoryReads"
+             description="The total number of GTI memory reads from Multisampling Color Cache (Multisampling Color Cache misses)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_reads"
+             units="messages"
+             symbol_name="GtiMscMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiVfMemoryReads"
+             description="The total number of GTI memory reads from Vertex Fetch."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="gti_vf_memory_reads"
+             units="messages"
+             symbol_name="GtiVfMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache (Stencil Cache misses)."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Reads $GtiL3Bank1Reads $GtiL3Bank2Reads $GtiL3Bank3Reads UADD UADD UADD"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA"
+                     availability="$SliceMask 0x01 AND $SkuRevisionId 0x02 ULT &amp;&amp;"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x13946000" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x0F968000" />
+        <register type="NOA" address="0x00009888" value="0x1196C000" />
+        <register type="NOA" address="0x00009888" value="0x13964000" />
+        <register type="NOA" address="0x00009888" value="0x11938000" />
+        <register type="NOA" address="0x00009888" value="0x1B93FE00" />
+        <register type="NOA" address="0x00009888" value="0x01940010" />
+        <register type="NOA" address="0x00009888" value="0x07941100" />
+        <register type="NOA" address="0x00009888" value="0x09941312" />
+        <register type="NOA" address="0x00009888" value="0x0B941514" />
+        <register type="NOA" address="0x00009888" value="0x0D941716" />
+        <register type="NOA" address="0x00009888" value="0x11940000" />
+        <register type="NOA" address="0x00009888" value="0x19940000" />
+        <register type="NOA" address="0x00009888" value="0x1B940000" />
+        <register type="NOA" address="0x00009888" value="0x1D940000" />
+        <register type="NOA" address="0x00009888" value="0x1B954000" />
+        <register type="NOA" address="0x00009888" value="0x1D95A550" />
+        <register type="NOA" address="0x00009888" value="0x1F9502AA" />
+        <register type="NOA" address="0x00009888" value="0x2F900157" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x13908000" />
+        <register type="NOA" address="0x00009888" value="0x21908000" />
+        <register type="NOA" address="0x00009888" value="0x23908000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27908000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00000D28" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x05 ULT $SkuRevisionId 0x02 UGTE &amp;&amp;"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x13946000" />
+        <register type="NOA" address="0x00009888" value="0x15940016" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x19930800" />
+        <register type="NOA" address="0x00009888" value="0x1B93AA55" />
+        <register type="NOA" address="0x00009888" value="0x1D9300AA" />
+        <register type="NOA" address="0x00009888" value="0x01940010" />
+        <register type="NOA" address="0x00009888" value="0x07941100" />
+        <register type="NOA" address="0x00009888" value="0x09941312" />
+        <register type="NOA" address="0x00009888" value="0x0B941514" />
+        <register type="NOA" address="0x00009888" value="0x0D941716" />
+        <register type="NOA" address="0x00009888" value="0x0F940018" />
+        <register type="NOA" address="0x00009888" value="0x1B940000" />
+        <register type="NOA" address="0x00009888" value="0x11940000" />
+        <register type="NOA" address="0x00009888" value="0x01E58000" />
+        <register type="NOA" address="0x00009888" value="0x03E57000" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x13908000" />
+        <register type="NOA" address="0x00009888" value="0x21908000" />
+        <register type="NOA" address="0x00009888" value="0x23908000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27908000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x2F908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C20" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+        <register type="NOA" address="0x00009888" value="0x47900421" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900421" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900061" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x05 UGTE"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900064" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900150" />
+        <register type="NOA" address="0x00009888" value="0x21900151" />
+        <register type="NOA" address="0x00009888" value="0x23900152" />
+        <register type="NOA" address="0x00009888" value="0x25900153" />
+        <register type="NOA" address="0x00009888" value="0x27900154" />
+        <register type="NOA" address="0x00009888" value="0x29900155" />
+        <register type="NOA" address="0x00009888" value="0x2B900156" />
+        <register type="NOA" address="0x00009888" value="0x2D900157" />
+        <register type="NOA" address="0x00009888" value="0x2F90015F" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F872" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="2ea0da8f-3527-4669-9d9d-13099a7435bf"
+       chipset="SKLGT2"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiMemoryWrites"
+             description="The total number of GTI memory writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all GTI accesses to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiMscMemoryWrites"
+             description="The total number of GTI memory writes from Multisampling Color Cache (Multisampling Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_writes"
+             units="messages"
+             symbol_name="GtiMscMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiCmdStreamerMemoryWrites"
+             description="The total number of GTI memory writes from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_writes"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiL3Bank0Writes"
+             description="The total number of GTI memory writes from L3 Bank 0 (L3 Bank 0 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_writes"
+             units="messages"
+             symbol_name="GtiL3Bank0Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank1Writes"
+             description="The total number of GTI memory writes from L3 Bank 1 (L3 Bank 1 invalidations)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_writes"
+             units="messages"
+             symbol_name="GtiL3Bank1Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank2Writes"
+             description="The total number of GTI memory writes from L3 Bank 2 (L3 Bank 2 invalidations)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_writes"
+             units="messages"
+             symbol_name="GtiL3Bank2Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank3Writes"
+             description="The total number of GTI memory writes from L3 Bank 3 (L3 Bank 3 invalidations)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_writes"
+             units="messages"
+             symbol_name="GtiL3Bank3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Writes $GtiL3Bank1Writes $GtiL3Bank2Writes $GtiL3Bank3Writes UADD UADD UADD"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRccMemoryWrites"
+             description="The total number of GTI memory writes from Render Color Cache (Render Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_writes"
+             units="messages"
+             symbol_name="GtiRccMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiSoMemoryWrites"
+             description="The total number of GTI memory writes from Stream Output."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_so_memory_writes"
+             units="messages"
+             symbol_name="GtiSoMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiStcMemoryWrites"
+             description="The total number of GTI memory writes from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_writes"
+             units="messages"
+             symbol_name="GtiStcMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GtiRczMemoryWrites"
+             description="The total number of GTI memory writes from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_writes"
+             units="messages"
+             symbol_name="GtiRczMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiHizMemoryWrites"
+             description="The total number of GTI memory writes from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_writes"
+             units="messages"
+             symbol_name="GtiHizMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA"
+                     availability="$SliceMask 0x01 AND $SkuRevisionId 0x02 ULT &amp;&amp;"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x13945400" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901400" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x0F968000" />
+        <register type="NOA" address="0x00009888" value="0x1196C000" />
+        <register type="NOA" address="0x00009888" value="0x13964000" />
+        <register type="NOA" address="0x00009888" value="0x11938000" />
+        <register type="NOA" address="0x00009888" value="0x1B93FE00" />
+        <register type="NOA" address="0x00009888" value="0x01940010" />
+        <register type="NOA" address="0x00009888" value="0x07941100" />
+        <register type="NOA" address="0x00009888" value="0x09941312" />
+        <register type="NOA" address="0x00009888" value="0x0B941514" />
+        <register type="NOA" address="0x00009888" value="0x0D941716" />
+        <register type="NOA" address="0x00009888" value="0x11940000" />
+        <register type="NOA" address="0x00009888" value="0x19940000" />
+        <register type="NOA" address="0x00009888" value="0x1B940000" />
+        <register type="NOA" address="0x00009888" value="0x1D940000" />
+        <register type="NOA" address="0x00009888" value="0x1B954000" />
+        <register type="NOA" address="0x00009888" value="0x1D95A550" />
+        <register type="NOA" address="0x00009888" value="0x1F9502AA" />
+        <register type="NOA" address="0x00009888" value="0x2F900167" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x13908000" />
+        <register type="NOA" address="0x00009888" value="0x21908000" />
+        <register type="NOA" address="0x00009888" value="0x23908000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27908000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00000D28" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x05 ULT $SkuRevisionId 0x02 UGTE &amp;&amp;"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x13945400" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901400" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x19930800" />
+        <register type="NOA" address="0x00009888" value="0x1B93AA55" />
+        <register type="NOA" address="0x00009888" value="0x1D93002A" />
+        <register type="NOA" address="0x00009888" value="0x01940010" />
+        <register type="NOA" address="0x00009888" value="0x07941100" />
+        <register type="NOA" address="0x00009888" value="0x09941312" />
+        <register type="NOA" address="0x00009888" value="0x0B941514" />
+        <register type="NOA" address="0x00009888" value="0x0D941716" />
+        <register type="NOA" address="0x00009888" value="0x1B940000" />
+        <register type="NOA" address="0x00009888" value="0x11940000" />
+        <register type="NOA" address="0x00009888" value="0x01E58000" />
+        <register type="NOA" address="0x00009888" value="0x03E57000" />
+        <register type="NOA" address="0x00009888" value="0x2F900167" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x13908000" />
+        <register type="NOA" address="0x00009888" value="0x21908000" />
+        <register type="NOA" address="0x00009888" value="0x23908000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27908000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C20" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+        <register type="NOA" address="0x00009888" value="0x47900421" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900421" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="NOA"
+                     availability="$SkuRevisionId 0x05 UGTE"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901000" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900160" />
+        <register type="NOA" address="0x00009888" value="0x21900161" />
+        <register type="NOA" address="0x00009888" value="0x23900162" />
+        <register type="NOA" address="0x00009888" value="0x25900163" />
+        <register type="NOA" address="0x00009888" value="0x27900164" />
+        <register type="NOA" address="0x00009888" value="0x29900165" />
+        <register type="NOA" address="0x00009888" value="0x2B900166" />
+        <register type="NOA" address="0x00009888" value="0x2D900167" />
+        <register type="NOA" address="0x00009888" value="0x2F900150" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F822" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="d97d16af-028b-4cd1-a672-6210cb5513dd"
+       chipset="SKLGT2"
+       symbol_name="ComputeExtended"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Writes 0"
+             description="The subslice 0 typed writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="typed_writes0"
+             units="messages"
+             symbol_name="TypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuTypedAtomics0"
+             description="The subslice 0 EU Typed Atomics subslice 0."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_typed_atomics0"
+             units="messages"
+             symbol_name="EuTypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Atomics 0"
+             description="The subslice 0 typed atomics."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="typed_atomics0"
+             units="messages"
+             symbol_name="TypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedAtomicsPerCacheLine"
+             description="The ratio of EU typed atomics requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedAtomics0 $TypedAtomics0 FDIV"
+             underscore_name="typed_atomics_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedAtomicsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedReads0"
+             description="The subslice 0 EU Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_untyped_reads0"
+             units="messages"
+             symbol_name="EuUntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Writes 0"
+             description="The subslice 0 untyped writes (including SLM writes)."
+             data_type="uint64"
+             equation="C 1 READ"
+             underscore_name="untyped_writes0"
+             units="messages"
+             symbol_name="UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedAtomics0"
+             description="The subslice 0 EU Untyped Atomics subslice 0."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_untyped_atomics0"
+             units="messages"
+             symbol_name="EuUntypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedWrites0"
+             description="The subslice 0 EU Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="eu_untyped_writes0"
+             units="messages"
+             symbol_name="EuUntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedWrites0"
+             description="The subslice 0 EU A64 Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_a64_untyped_writes0"
+             units="messages"
+             symbol_name="EuA64UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedWritesPerCacheLine"
+             description="The ratio of EU untyped write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuUntypedWrites0 $EuA64UntypedWrites0 UADD $UntypedWrites0 FDIV"
+             underscore_name="untyped_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedWrites0"
+             description="The subslice 0 EU Typed Writes subslice 0."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="eu_typed_writes0"
+             units="messages"
+             symbol_name="EuTypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedWritesPerCacheLine"
+             description="The ratio of EU typed write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedWrites0 $TypedWrites0 FDIV"
+             underscore_name="typed_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Reads 0"
+             description="The subslice 0 typed reads."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="typed_reads0"
+             units="messages"
+             symbol_name="TypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Reads 0"
+             description="The subslice 0 untyped reads (including SLM reads)."
+             data_type="uint64"
+             equation="C 3 READ"
+             underscore_name="untyped_reads0"
+             units="messages"
+             symbol_name="UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedReads0"
+             description="The subslice 0 EU A64 Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_a64_untyped_reads0"
+             units="messages"
+             symbol_name="EuA64UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedReads0"
+             description="The subslice 0 EU Typed Reads subslice 0."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="eu_typed_reads0"
+             units="messages"
+             symbol_name="EuTypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedReadsPerCacheLine"
+             description="The ratio of EU untyped read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuUntypedReads0 $EuA64UntypedReads0 UADD $UntypedReads0 FDIV"
+             underscore_name="untyped_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="The ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuTypedReads0 $TypedReads0 FDIV"
+             underscore_name="typed_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA"
+                     availability="$SubsliceMask 0x01 AND"
+                     priority="0"
+                     >
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x141C8160" />
+        <register type="NOA" address="0x00009888" value="0x161C8015" />
+        <register type="NOA" address="0x00009888" value="0x181C0120" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4EAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B01" />
+        <register type="NOA" address="0x00009888" value="0x006C0200" />
+        <register type="NOA" address="0x00009888" value="0x026C000C" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x001C0041" />
+        <register type="NOA" address="0x00009888" value="0x061C4200" />
+        <register type="NOA" address="0x00009888" value="0x081C4443" />
+        <register type="NOA" address="0x00009888" value="0x0A1C4645" />
+        <register type="NOA" address="0x00009888" value="0x0C1C7647" />
+        <register type="NOA" address="0x00009888" value="0x041C7357" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x101C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0000" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4CAA2A" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02AA" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5515" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00000D28" value="0x00000000" />
+        <register type="NOA" address="0x00009888" value="0x11907FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900040" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900802" />
+        <register type="NOA" address="0x00009888" value="0x47900842" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900842" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900800" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FC2A" />
+        <register type="OA" address="0x00002774" value="0x0000BF00" />
+        <register type="OA" address="0x00002778" value="0x0007FC6A" />
+        <register type="OA" address="0x0000277C" value="0x0000BF00" />
+        <register type="OA" address="0x00002780" value="0x0007FC92" />
+        <register type="OA" address="0x00002784" value="0x0000BF00" />
+        <register type="OA" address="0x00002788" value="0x0007FCA2" />
+        <register type="OA" address="0x0000278C" value="0x0000BF00" />
+        <register type="OA" address="0x00002790" value="0x0007FC32" />
+        <register type="OA" address="0x00002794" value="0x0000BF00" />
+        <register type="OA" address="0x00002798" value="0x0007FC9A" />
+        <register type="OA" address="0x0000279C" value="0x0000BF00" />
+        <register type="OA" address="0x000027A0" value="0x0007FE6A" />
+        <register type="OA" address="0x000027A4" value="0x0000BF00" />
+        <register type="OA" address="0x000027A8" value="0x0007FE7A" />
+        <register type="OA" address="0x000027AC" value="0x0000BF00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics L3 Cache Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_l3_cache"
+       hw_config_guid="9fb22842-e708-43f7-9752-e0e41670c39e"
+       chipset="SKLGT2"
+       symbol_name="ComputeL3Cache"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 03 Accesses"
+             description="The total number of accesses to L3 Bank 03."
+             data_type="uint64"
+             equation="B 3 READ 2 UMUL"
+             underscore_name="l3_bank03_accesses"
+             units="messages"
+             symbol_name="L3Bank03Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Accesses"
+             description="The total number of L3 accesses from all entities."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ B 2 READ B 3 READ UADD UADD UADD 2 UMUL"
+             underscore_name="l3_accesses"
+             units="messages"
+             symbol_name="L3Accesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 29 READ 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="EU FPU0 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu0_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ C 5 READ UADD"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="L3 Bank 00 Accesses"
+             description="The total number of accesses to L3 Bank 00."
+             data_type="uint64"
+             equation="C 0 READ 2 UMUL"
+             underscore_name="l3_bank00_accesses"
+             units="messages"
+             symbol_name="L3Bank00Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU0 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 19 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu0_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 14 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu1_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 00 IC Accesses"
+             description="The total number of accesses to L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 0 READ B 1 READ UADD 2 UMUL $L3Bank00Accesses UMIN"
+             underscore_name="l3_bank00_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank00IcAccesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 00 IC Hits"
+             description="The total number of hits in L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 1 READ 2 UMUL $L3Bank00IcAccesses UMIN"
+             underscore_name="l3_bank00_ic_hits"
+             units="messages"
+             symbol_name="L3Bank00IcHits"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="L3 Bank 01 Accesses"
+             description="The total number of accesses to L3 Bank 01."
+             data_type="uint64"
+             equation="C 1 READ 2 UMUL"
+             underscore_name="l3_bank01_accesses"
+             units="messages"
+             symbol_name="L3Bank01Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 20 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu1_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="L3 Bank 02 Accesses"
+             description="The total number of accesses to L3 Bank 02."
+             data_type="uint64"
+             equation="B 2 READ 2 UMUL"
+             underscore_name="l3_bank02_accesses"
+             units="messages"
+             symbol_name="L3Bank02Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="L3 Total Throughput"
+             description="The total number of GPU memory bytes transferred via L3."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Accesses 64 UMUL"
+             underscore_name="l3_total_throughput"
+             units="bytes"
+             symbol_name="L3TotalThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="EU FPU1 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C0760" />
+        <register type="NOA" address="0x00009888" value="0x1593001E" />
+        <register type="NOA" address="0x00009888" value="0x3F901403" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E8020" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x006C0051" />
+        <register type="NOA" address="0x00009888" value="0x066C5000" />
+        <register type="NOA" address="0x00009888" value="0x086C5C5D" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5E5F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x061B8000" />
+        <register type="NOA" address="0x00009888" value="0x081BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1CE000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2A00" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0280" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F1500" />
+        <register type="NOA" address="0x00009888" value="0x100F0140" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C0A00" />
+        <register type="NOA" address="0x00009888" value="0x182C00A0" />
+        <register type="NOA" address="0x00009888" value="0x03933300" />
+        <register type="NOA" address="0x00009888" value="0x05930032" />
+        <register type="NOA" address="0x00009888" value="0x11930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900167" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190030F" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900042" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x53901111" />
+        <register type="NOA" address="0x00009888" value="0x43900420" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFFA" />
+        <register type="OA" address="0x00002774" value="0x0000FEFE" />
+        <register type="OA" address="0x00002778" value="0x0007FFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000FEFD" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000FBEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000FBDF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00101100" />
+        <register type="FLEX" address="0x0000E45C" value="0x00201200" />
+        <register type="FLEX" address="0x0000E55C" value="0x00301300" />
+        <register type="FLEX" address="0x0000E65C" value="0x00401400" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="5378e2a1-4248-4188-a4ae-da25a794c603"
+       chipset="SKLGT2"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ C 2 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader02_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader02AccessStalledOnL3"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F0232" />
+        <register type="NOA" address="0x00009888" value="0x124F4640" />
+        <register type="NOA" address="0x00009888" value="0x106C0232" />
+        <register type="NOA" address="0x00009888" value="0x11834400" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x004F1880" />
+        <register type="NOA" address="0x00009888" value="0x024F08BB" />
+        <register type="NOA" address="0x00009888" value="0x044F001B" />
+        <register type="NOA" address="0x00009888" value="0x046C0100" />
+        <register type="NOA" address="0x00009888" value="0x066C000B" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x041B8000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025BC000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x165C8000" />
+        <register type="NOA" address="0x00009888" value="0x185C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00A0" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x062CC000" />
+        <register type="NOA" address="0x00009888" value="0x082CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x0F8305C0" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x07830000" />
+        <register type="NOA" address="0x00009888" value="0x1D950080" />
+        <register type="NOA" address="0x00009888" value="0x13928000" />
+        <register type="NOA" address="0x00009888" value="0x0F988000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x4B9000A0" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="f42cdd6a-b000-42cb-870f-5eb423a7f514"
+       chipset="SKLGT2"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C7B40" />
+        <register type="NOA" address="0x00009888" value="0x166C0020" />
+        <register type="NOA" address="0x00009888" value="0x0A603444" />
+        <register type="NOA" address="0x00009888" value="0x0A613400" />
+        <register type="NOA" address="0x00009888" value="0x1A4EA800" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C5327" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5425" />
+        <register type="NOA" address="0x00009888" value="0x006C2A00" />
+        <register type="NOA" address="0x00009888" value="0x026C285B" />
+        <register type="NOA" address="0x00009888" value="0x046C005C" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0800" />
+        <register type="NOA" address="0x00009888" value="0x0C1BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C003C" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x10600000" />
+        <register type="NOA" address="0x00009888" value="0x04600000" />
+        <register type="NOA" address="0x00009888" value="0x0C610044" />
+        <register type="NOA" address="0x00009888" value="0x10610000" />
+        <register type="NOA" address="0x00009888" value="0x06610000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02A8" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0154" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190FFC0" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900420" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900021" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900400" />
+        <register type="NOA" address="0x00009888" value="0x43900421" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_2"
+       hw_config_guid="b9bf2423-d88c-4a7b-a051-627611d00dcc"
+       chipset="SKLGT2"
+       symbol_name="L3_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Stalled"
+             description="The percentage of time in which slice0 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_stalled"
+             units="percent"
+             symbol_name="L30Bank2Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Active"
+             description="The percentage of time in which slice0 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_active"
+             units="percent"
+             symbol_name="L30Bank2Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C02E0" />
+        <register type="NOA" address="0x00009888" value="0x146C0001" />
+        <register type="NOA" address="0x00009888" value="0x0A623400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x026C3324" />
+        <register type="NOA" address="0x00009888" value="0x046C3422" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x06614000" />
+        <register type="NOA" address="0x00009888" value="0x0C620044" />
+        <register type="NOA" address="0x00009888" value="0x10620000" />
+        <register type="NOA" address="0x00009888" value="0x06620000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_3"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_3"
+       hw_config_guid="2414a93d-d84f-406e-99c0-472161194b40"
+       chipset="SKLGT2"
+       symbol_name="L3_3"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 L3 Bank3 Stalled"
+             description="The percentage of time in which slice0 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_stalled"
+             units="percent"
+             symbol_name="L30Bank3Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank3 Active"
+             description="The percentage of time in which slice0 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_active"
+             units="percent"
+             symbol_name="L30Bank3Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C4E80" />
+        <register type="NOA" address="0x00009888" value="0x146C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A633400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x026C3321" />
+        <register type="NOA" address="0x00009888" value="0x046C342F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C2000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x06604000" />
+        <register type="NOA" address="0x00009888" value="0x0C630044" />
+        <register type="NOA" address="0x00009888" value="0x10630000" />
+        <register type="NOA" address="0x00009888" value="0x06630000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00AA" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900002" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="53a45d2d-170b-4cf5-b7bb-585120c8e2f5"
+       chipset="SKLGT2"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x102F3800" />
+        <register type="NOA" address="0x00009888" value="0x144D0500" />
+        <register type="NOA" address="0x00009888" value="0x120D03C0" />
+        <register type="NOA" address="0x00009888" value="0x140D03CF" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0004" />
+        <register type="NOA" address="0x00009888" value="0x0C4E4000" />
+        <register type="NOA" address="0x00009888" value="0x042F0480" />
+        <register type="NOA" address="0x00009888" value="0x082F0000" />
+        <register type="NOA" address="0x00009888" value="0x022F0000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0090" />
+        <register type="NOA" address="0x00009888" value="0x064D0027" />
+        <register type="NOA" address="0x00009888" value="0x004D0000" />
+        <register type="NOA" address="0x00009888" value="0x000D0D40" />
+        <register type="NOA" address="0x00009888" value="0x020D803F" />
+        <register type="NOA" address="0x00009888" value="0x040D8023" />
+        <register type="NOA" address="0x00009888" value="0x100D0000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020F0010" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0050" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41901400" />
+        <register type="NOA" address="0x00009888" value="0x43901485" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900001" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000EFFF" />
+        <register type="OA" address="0x00002778" value="0x00006000" />
+        <register type="OA" address="0x0000277C" value="0x0000F3FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler"
+       hw_config_guid="b4cff514-a91e-4798-a0b3-426ca13fc9c1"
+       chipset="SKLGT2"
+       symbol_name="Sampler"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x14152C00" />
+        <register type="NOA" address="0x00009888" value="0x16150005" />
+        <register type="NOA" address="0x00009888" value="0x121600A0" />
+        <register type="NOA" address="0x00009888" value="0x14352C00" />
+        <register type="NOA" address="0x00009888" value="0x16350005" />
+        <register type="NOA" address="0x00009888" value="0x123600A0" />
+        <register type="NOA" address="0x00009888" value="0x14552C00" />
+        <register type="NOA" address="0x00009888" value="0x16550005" />
+        <register type="NOA" address="0x00009888" value="0x125600A0" />
+        <register type="NOA" address="0x00009888" value="0x062F6000" />
+        <register type="NOA" address="0x00009888" value="0x022F2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0050" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0010" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0350" />
+        <register type="NOA" address="0x00009888" value="0x0C0FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F00DA" />
+        <register type="NOA" address="0x00009888" value="0x182C0028" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x022DC000" />
+        <register type="NOA" address="0x00009888" value="0x042D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C138000" />
+        <register type="NOA" address="0x00009888" value="0x0E132000" />
+        <register type="NOA" address="0x00009888" value="0x0413C000" />
+        <register type="NOA" address="0x00009888" value="0x1C140018" />
+        <register type="NOA" address="0x00009888" value="0x0C157000" />
+        <register type="NOA" address="0x00009888" value="0x0E150078" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x04162180" />
+        <register type="NOA" address="0x00009888" value="0x02160000" />
+        <register type="NOA" address="0x00009888" value="0x04174000" />
+        <register type="NOA" address="0x00009888" value="0x0233A000" />
+        <register type="NOA" address="0x00009888" value="0x04333000" />
+        <register type="NOA" address="0x00009888" value="0x14348000" />
+        <register type="NOA" address="0x00009888" value="0x16348000" />
+        <register type="NOA" address="0x00009888" value="0x02357870" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x04360043" />
+        <register type="NOA" address="0x00009888" value="0x02360000" />
+        <register type="NOA" address="0x00009888" value="0x04371000" />
+        <register type="NOA" address="0x00009888" value="0x0E538000" />
+        <register type="NOA" address="0x00009888" value="0x00538000" />
+        <register type="NOA" address="0x00009888" value="0x06533000" />
+        <register type="NOA" address="0x00009888" value="0x1C540020" />
+        <register type="NOA" address="0x00009888" value="0x12548000" />
+        <register type="NOA" address="0x00009888" value="0x0E557000" />
+        <register type="NOA" address="0x00009888" value="0x00557800" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x06560043" />
+        <register type="NOA" address="0x00009888" value="0x02560000" />
+        <register type="NOA" address="0x00009888" value="0x06571000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900060" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="7821d13b-9b8b-4405-9618-78cd56b62cce"
+       chipset="SKLGT2"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12120000" />
+        <register type="NOA" address="0x00009888" value="0x12320000" />
+        <register type="NOA" address="0x00009888" value="0x12520000" />
+        <register type="NOA" address="0x00009888" value="0x002F8000" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0015" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F03A0" />
+        <register type="NOA" address="0x00009888" value="0x0C0FF000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0095" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D4000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x02108000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x02118000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x02121880" />
+        <register type="NOA" address="0x00009888" value="0x041219B5" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x02134000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x0C308000" />
+        <register type="NOA" address="0x00009888" value="0x0E304000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x0C318000" />
+        <register type="NOA" address="0x00009888" value="0x0E314000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x0C321A80" />
+        <register type="NOA" address="0x00009888" value="0x0E320033" />
+        <register type="NOA" address="0x00009888" value="0x06320031" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x0C334000" />
+        <register type="NOA" address="0x00009888" value="0x0E331000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0E508000" />
+        <register type="NOA" address="0x00009888" value="0x00508000" />
+        <register type="NOA" address="0x00009888" value="0x02504000" />
+        <register type="NOA" address="0x00009888" value="0x0E518000" />
+        <register type="NOA" address="0x00009888" value="0x00518000" />
+        <register type="NOA" address="0x00009888" value="0x02514000" />
+        <register type="NOA" address="0x00009888" value="0x0E521880" />
+        <register type="NOA" address="0x00009888" value="0x00521A80" />
+        <register type="NOA" address="0x00009888" value="0x02520033" />
+        <register type="NOA" address="0x00009888" value="0x0E534000" />
+        <register type="NOA" address="0x00009888" value="0x00534000" />
+        <register type="NOA" address="0x00009888" value="0x02531000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900062" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x00007FFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x00009FFF" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000EFFF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000F3FF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FDFF" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="893f1a4d-919d-4388-8cb7-746d73ea7259"
+       chipset="SKLGT2"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12124D60" />
+        <register type="NOA" address="0x00009888" value="0x12322E60" />
+        <register type="NOA" address="0x00009888" value="0x12524D60" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0014" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0FE000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0097" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x002D8000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x04121FB7" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x00308000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x00318000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x00321B80" />
+        <register type="NOA" address="0x00009888" value="0x0632003F" />
+        <register type="NOA" address="0x00009888" value="0x00334000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0250C000" />
+        <register type="NOA" address="0x00009888" value="0x0251C000" />
+        <register type="NOA" address="0x00009888" value="0x02521FB7" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x02535000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900063" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extra Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extra"
+       hw_config_guid="41a24047-7484-4ead-ae37-de907e5ff2b2"
+       chipset="SKLGT2"
+       symbol_name="ComputeExtra"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active including Ext Math"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing including Extended Math processing"
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ  C 7 READ C 6 READ FADD C 5 READ FADD 8 FMUL FADD 100 FMUL $EuCoresTotalCount FDIV $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active_adjusted"
+             units="percent"
+             symbol_name="Fpu1ActiveAdjusted"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121203E0" />
+        <register type="NOA" address="0x00009888" value="0x123203E0" />
+        <register type="NOA" address="0x00009888" value="0x125203E0" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F006C" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x042D8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06114000" />
+        <register type="NOA" address="0x00009888" value="0x06120033" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x04308000" />
+        <register type="NOA" address="0x00009888" value="0x04318000" />
+        <register type="NOA" address="0x00009888" value="0x04321980" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x04334000" />
+        <register type="NOA" address="0x00009888" value="0x04504000" />
+        <register type="NOA" address="0x00009888" value="0x04514000" />
+        <register type="NOA" address="0x00009888" value="0x04520033" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x04531000" />
+        <register type="NOA" address="0x00009888" value="0x1190E000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x43900C00" />
+        <register type="NOA" address="0x00009888" value="0x45900002" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00001000" />
+        <register type="FLEX" address="0x0000E558" value="0x00003002" />
+        <register type="FLEX" address="0x0000E658" value="0x00005004" />
+        <register type="FLEX" address="0x0000E758" value="0x00011010" />
+        <register type="FLEX" address="0x0000E45C" value="0x00050012" />
+        <register type="FLEX" address="0x0000E55C" value="0x00052051" />
+        <register type="FLEX" address="0x0000E65C" value="0x00000008" />
+    </register_config>
+  </set>
+
+  <set name="Media Vme Pipe Gen9"
+       mdapi_supported_apis="MEDIA IO BB"
+       underscore_name="vme_pipe"
+       hw_config_guid="95910492-943f-44bd-9461-390240f243fd"
+       chipset="SKLGT2"
+       symbol_name="VMEPipe"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL  GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="8 A 10 READ FMUL $EuThreadsCount FDIV $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VME Busy"
+             description="The percentage of time in which VME (IME or CRE) was actively processing data."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ B 3 READ FADD 2 FDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vme_busy"
+             units="percent"
+             symbol_name="VMEBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Overview System Batch Tier2"
+             mdapi_group="VME Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A5800" />
+        <register type="NOA" address="0x00009888" value="0x161A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12180240" />
+        <register type="NOA" address="0x00009888" value="0x14180002" />
+        <register type="NOA" address="0x00009888" value="0x143A5800" />
+        <register type="NOA" address="0x00009888" value="0x163A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12380240" />
+        <register type="NOA" address="0x00009888" value="0x14380002" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x022F8000" />
+        <register type="NOA" address="0x00009888" value="0x042F3000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C1500" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F9500" />
+        <register type="NOA" address="0x00009888" value="0x100F002A" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C0A00" />
+        <register type="NOA" address="0x00009888" value="0x0A2DC000" />
+        <register type="NOA" address="0x00009888" value="0x0C2DC000" />
+        <register type="NOA" address="0x00009888" value="0x04193000" />
+        <register type="NOA" address="0x00009888" value="0x081A28C1" />
+        <register type="NOA" address="0x00009888" value="0x001A0000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x0613C000" />
+        <register type="NOA" address="0x00009888" value="0x0813F000" />
+        <register type="NOA" address="0x00009888" value="0x00172000" />
+        <register type="NOA" address="0x00009888" value="0x06178000" />
+        <register type="NOA" address="0x00009888" value="0x0817A000" />
+        <register type="NOA" address="0x00009888" value="0x00180037" />
+        <register type="NOA" address="0x00009888" value="0x06180940" />
+        <register type="NOA" address="0x00009888" value="0x08180000" />
+        <register type="NOA" address="0x00009888" value="0x02180000" />
+        <register type="NOA" address="0x00009888" value="0x04183000" />
+        <register type="NOA" address="0x00009888" value="0x06393000" />
+        <register type="NOA" address="0x00009888" value="0x0C3A28C1" />
+        <register type="NOA" address="0x00009888" value="0x003A0000" />
+        <register type="NOA" address="0x00009888" value="0x0A33F000" />
+        <register type="NOA" address="0x00009888" value="0x0C33F000" />
+        <register type="NOA" address="0x00009888" value="0x0A37A000" />
+        <register type="NOA" address="0x00009888" value="0x0C37A000" />
+        <register type="NOA" address="0x00009888" value="0x0A380977" />
+        <register type="NOA" address="0x00009888" value="0x08380000" />
+        <register type="NOA" address="0x00009888" value="0x04380000" />
+        <register type="NOA" address="0x00009888" value="0x06383000" />
+        <register type="NOA" address="0x00009888" value="0x119000FF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900040" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900800" />
+        <register type="NOA" address="0x00009888" value="0x47901000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900844" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00100030" />
+        <register type="OA" address="0x00002774" value="0x0000FFF9" />
+        <register type="OA" address="0x00002778" value="0x00000002" />
+        <register type="OA" address="0x0000277C" value="0x0000FFFC" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFF3" />
+        <register type="OA" address="0x00002788" value="0x00100180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FF3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00008003" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="1651949f-0ac0-4cb1-a06f-dafd74a407d1"
+       chipset="SKLGT2"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810000" />
+        <register type="NOA" address="0x00009888" value="0x07810016" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930040" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
diff --git a/src/mesa/drivers/dri/i965/brw_oa_sklgt3.xml b/src/mesa/drivers/dri/i965/brw_oa_sklgt3.xml
new file mode 100644
index 0000000..6dd80d6
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_sklgt3.xml
@@ -0,0 +1,10499 @@
+<?xml version="1.0"?>
+<metrics version="1491577975" merge_md5="">
+  <set name="Render Metrics Basic Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="4616d450-2393-4836-8146-53c5ed84d359"
+       chipset="SKLGT3"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Cache Misses"
+             description="The total number of sampler cache misses in all LODs in all sampler units."
+             data_type="uint64"
+             equation="B 4 READ B 5 READ UADD 8 UMUL"
+             underscore_name="sampler_l1_misses"
+             units="messages"
+             symbol_name="SamplerL1Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler 1 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 1 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler1_bottleneck"
+             units="percent"
+             symbol_name="Sampler1Bottleneck"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$SamplerL1Misses 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="$Sampler0Busy $Sampler1Busy FMAX"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="L3 Lookup Accesses w/o IC"
+             description="The total number of L3 cache lookup accesses w/o IC."
+             data_type="uint64"
+             equation="$SamplerL1Misses $ShaderMemoryAccesses UADD"
+             underscore_name="l3_lookups"
+             units="messages"
+             symbol_name="L3Lookups"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck $Sampler1Bottleneck FMAX"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C01E0" />
+        <register type="NOA" address="0x00009888" value="0x12170280" />
+        <register type="NOA" address="0x00009888" value="0x12370280" />
+        <register type="NOA" address="0x00009888" value="0x16EC01E0" />
+        <register type="NOA" address="0x00009888" value="0x11930317" />
+        <register type="NOA" address="0x00009888" value="0x159303DF" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0380" />
+        <register type="NOA" address="0x00009888" value="0x0A6C0053" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A1B4000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0001" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x042F1000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C8400" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0002" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F6600" />
+        <register type="NOA" address="0x00009888" value="0x100F0001" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CA200" />
+        <register type="NOA" address="0x00009888" value="0x062D8000" />
+        <register type="NOA" address="0x00009888" value="0x082D8000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x08133000" />
+        <register type="NOA" address="0x00009888" value="0x00170020" />
+        <register type="NOA" address="0x00009888" value="0x08170021" />
+        <register type="NOA" address="0x00009888" value="0x10170000" />
+        <register type="NOA" address="0x00009888" value="0x0633C000" />
+        <register type="NOA" address="0x00009888" value="0x0833C000" />
+        <register type="NOA" address="0x00009888" value="0x06370800" />
+        <register type="NOA" address="0x00009888" value="0x08370840" />
+        <register type="NOA" address="0x00009888" value="0x10370000" />
+        <register type="NOA" address="0x00009888" value="0x1ACE0200" />
+        <register type="NOA" address="0x00009888" value="0x0AEC5300" />
+        <register type="NOA" address="0x00009888" value="0x10EC0000" />
+        <register type="NOA" address="0x00009888" value="0x1CEC0000" />
+        <register type="NOA" address="0x00009888" value="0x0A9B8000" />
+        <register type="NOA" address="0x00009888" value="0x1C9C0002" />
+        <register type="NOA" address="0x00009888" value="0x0CCC0002" />
+        <register type="NOA" address="0x00009888" value="0x0A8D8000" />
+        <register type="NOA" address="0x00009888" value="0x108F0001" />
+        <register type="NOA" address="0x00009888" value="0x16AC8000" />
+        <register type="NOA" address="0x00009888" value="0x0D933031" />
+        <register type="NOA" address="0x00009888" value="0x0F933E3F" />
+        <register type="NOA" address="0x00009888" value="0x01933D00" />
+        <register type="NOA" address="0x00009888" value="0x0393073C" />
+        <register type="NOA" address="0x00009888" value="0x0593000E" />
+        <register type="NOA" address="0x00009888" value="0x1D930000" />
+        <register type="NOA" address="0x00009888" value="0x19930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x2F908000" />
+        <register type="NOA" address="0x00009888" value="0x31908000" />
+        <register type="NOA" address="0x00009888" value="0x15908000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190003F" />
+        <register type="NOA" address="0x00009888" value="0x51907710" />
+        <register type="NOA" address="0x00009888" value="0x419020A0" />
+        <register type="NOA" address="0x00009888" value="0x55901515" />
+        <register type="NOA" address="0x00009888" value="0x45900529" />
+        <register type="NOA" address="0x00009888" value="0x47901025" />
+        <register type="NOA" address="0x00009888" value="0x57907770" />
+        <register type="NOA" address="0x00009888" value="0x49902100" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900108" />
+        <register type="NOA" address="0x00009888" value="0x59900007" />
+        <register type="NOA" address="0x00009888" value="0x43902108" />
+        <register type="NOA" address="0x00009888" value="0x53907777" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="4320492b-fd03-42ac-922f-dbe1ef3b7b58"
+       chipset="SKLGT3"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ C 0 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 3 READ B 4 READ B 5 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 1 READ C 2 READ C 3 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ B 2 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x124F1C00" />
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0820" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x064F0900" />
+        <register type="NOA" address="0x00009888" value="0x084F0032" />
+        <register type="NOA" address="0x00009888" value="0x0A4F1891" />
+        <register type="NOA" address="0x00009888" value="0x0C4F0E00" />
+        <register type="NOA" address="0x00009888" value="0x0E4F003C" />
+        <register type="NOA" address="0x00009888" value="0x004F0D80" />
+        <register type="NOA" address="0x00009888" value="0x024F003B" />
+        <register type="NOA" address="0x00009888" value="0x006C0002" />
+        <register type="NOA" address="0x00009888" value="0x086C0100" />
+        <register type="NOA" address="0x00009888" value="0x0C6C000C" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B00" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x081B8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E1B8000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C8000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0024" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5BC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C6000" />
+        <register type="NOA" address="0x00009888" value="0x1C5C001B" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0208" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2CC000" />
+        <register type="NOA" address="0x00009888" value="0x162CFB00" />
+        <register type="NOA" address="0x00009888" value="0x182C00BE" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x19900157" />
+        <register type="NOA" address="0x00009888" value="0x1B900158" />
+        <register type="NOA" address="0x00009888" value="0x1D900105" />
+        <register type="NOA" address="0x00009888" value="0x1F900103" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x11900FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900863" />
+        <register type="NOA" address="0x00009888" value="0x47900802" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900802" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900002" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900C62" />
+        <register type="NOA" address="0x00009888" value="0x53903333" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="bd2d9cae-b9ec-4f5b-9d2f-934bed398a2d"
+       chipset="SKLGT3"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x0C0E001F" />
+        <register type="NOA" address="0x00009888" value="0x0A0F0000" />
+        <register type="NOA" address="0x00009888" value="0x10116800" />
+        <register type="NOA" address="0x00009888" value="0x178A03E0" />
+        <register type="NOA" address="0x00009888" value="0x11824C00" />
+        <register type="NOA" address="0x00009888" value="0x11830020" />
+        <register type="NOA" address="0x00009888" value="0x13840020" />
+        <register type="NOA" address="0x00009888" value="0x11850019" />
+        <register type="NOA" address="0x00009888" value="0x11860007" />
+        <register type="NOA" address="0x00009888" value="0x01870C40" />
+        <register type="NOA" address="0x00009888" value="0x17880000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x040D4000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020E5400" />
+        <register type="NOA" address="0x00009888" value="0x000E0000" />
+        <register type="NOA" address="0x00009888" value="0x080F0040" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x100F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0040" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06110012" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x01898000" />
+        <register type="NOA" address="0x00009888" value="0x0D890100" />
+        <register type="NOA" address="0x00009888" value="0x03898000" />
+        <register type="NOA" address="0x00009888" value="0x09808000" />
+        <register type="NOA" address="0x00009888" value="0x0B808000" />
+        <register type="NOA" address="0x00009888" value="0x0380C000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A0075" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0000" />
+        <register type="NOA" address="0x00009888" value="0x118A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A4000" />
+        <register type="NOA" address="0x00009888" value="0x138A8000" />
+        <register type="NOA" address="0x00009888" value="0x1D81A000" />
+        <register type="NOA" address="0x00009888" value="0x15818000" />
+        <register type="NOA" address="0x00009888" value="0x17818000" />
+        <register type="NOA" address="0x00009888" value="0x0B820030" />
+        <register type="NOA" address="0x00009888" value="0x07828000" />
+        <register type="NOA" address="0x00009888" value="0x0D824000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x05824000" />
+        <register type="NOA" address="0x00009888" value="0x0D830003" />
+        <register type="NOA" address="0x00009888" value="0x0583000C" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x03838000" />
+        <register type="NOA" address="0x00009888" value="0x07838000" />
+        <register type="NOA" address="0x00009888" value="0x0B840980" />
+        <register type="NOA" address="0x00009888" value="0x03844D80" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x09850080" />
+        <register type="NOA" address="0x00009888" value="0x03850003" />
+        <register type="NOA" address="0x00009888" value="0x01850000" />
+        <register type="NOA" address="0x00009888" value="0x07860000" />
+        <register type="NOA" address="0x00009888" value="0x0F860400" />
+        <register type="NOA" address="0x00009888" value="0x09870032" />
+        <register type="NOA" address="0x00009888" value="0x01888052" />
+        <register type="NOA" address="0x00009888" value="0x11880000" />
+        <register type="NOA" address="0x00009888" value="0x09884000" />
+        <register type="NOA" address="0x00009888" value="0x1B931001" />
+        <register type="NOA" address="0x00009888" value="0x1D930001" />
+        <register type="NOA" address="0x00009888" value="0x19934000" />
+        <register type="NOA" address="0x00009888" value="0x1B958000" />
+        <register type="NOA" address="0x00009888" value="0x1D950094" />
+        <register type="NOA" address="0x00009888" value="0x19958000" />
+        <register type="NOA" address="0x00009888" value="0x09E58000" />
+        <register type="NOA" address="0x00009888" value="0x0BE58000" />
+        <register type="NOA" address="0x00009888" value="0x03E5C000" />
+        <register type="NOA" address="0x00009888" value="0x0592C000" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D924000" />
+        <register type="NOA" address="0x00009888" value="0x0F924000" />
+        <register type="NOA" address="0x00009888" value="0x11928000" />
+        <register type="NOA" address="0x00009888" value="0x1392C000" />
+        <register type="NOA" address="0x00009888" value="0x09924000" />
+        <register type="NOA" address="0x00009888" value="0x01985000" />
+        <register type="NOA" address="0x00009888" value="0x07988000" />
+        <register type="NOA" address="0x00009888" value="0x09981000" />
+        <register type="NOA" address="0x00009888" value="0x0B982000" />
+        <register type="NOA" address="0x00009888" value="0x0D982000" />
+        <register type="NOA" address="0x00009888" value="0x0F989000" />
+        <register type="NOA" address="0x00009888" value="0x05982000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1190C080" />
+        <register type="NOA" address="0x00009888" value="0x51901150" />
+        <register type="NOA" address="0x00009888" value="0x41901400" />
+        <register type="NOA" address="0x00009888" value="0x55905111" />
+        <register type="NOA" address="0x00009888" value="0x45901400" />
+        <register type="NOA" address="0x00009888" value="0x479004A5" />
+        <register type="NOA" address="0x00009888" value="0x57903455" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B9000A0" />
+        <register type="NOA" address="0x00009888" value="0x59900001" />
+        <register type="NOA" address="0x00009888" value="0x43900005" />
+        <register type="NOA" address="0x00009888" value="0x53900455" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="4ca0f3fe-7fd3-4924-98cb-1807d9879767"
+       chipset="SKLGT3"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank0Reads"
+             description="The total number of GTI memory reads from L3 Bank 0 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_reads"
+             units="messages"
+             symbol_name="GtiL3Bank0Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all accesses from GTI to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiL3Bank3Reads"
+             description="The total number of GTI memory reads from L3 Bank 3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_reads"
+             units="messages"
+             symbol_name="GtiL3Bank3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiRsMemoryReads"
+             description="The total number of GTI memory reads from Resource Streamer."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_rs_memory_reads"
+             units="messages"
+             symbol_name="GtiRsMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiHizMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache (Hi-Depth Cache misses)."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_reads"
+             units="messages"
+             symbol_name="GtiHizMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="GtiRccMemoryReads"
+             description="The total number of GTI memory reads from Render Color Cache (Render Color Cache misses)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_reads"
+             units="messages"
+             symbol_name="GtiRccMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank1Reads"
+             description="The total number of GTI memory reads from L3 Bank 1 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_reads"
+             units="messages"
+             symbol_name="GtiL3Bank1Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiCmdStreamerMemoryReads"
+             description="The total number of GTI memory reads from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_reads"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GtiL3Bank2Reads"
+             description="The total number of GTI memory reads from L3 Bank 2 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_reads"
+             units="messages"
+             symbol_name="GtiL3Bank2Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiMemoryReads"
+             description="The total number of GTI memory reads."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache (Render Depth Cache misses)."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="GtiMscMemoryReads"
+             description="The total number of GTI memory reads from Multisampling Color Cache (Multisampling Color Cache misses)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_reads"
+             units="messages"
+             symbol_name="GtiMscMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiVfMemoryReads"
+             description="The total number of GTI memory reads from Vertex Fetch."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="gti_vf_memory_reads"
+             units="messages"
+             symbol_name="GtiVfMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache (Stencil Cache misses)."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Reads $GtiL3Bank1Reads $GtiL3Bank2Reads $GtiL3Bank3Reads UADD UADD UADD"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900064" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900150" />
+        <register type="NOA" address="0x00009888" value="0x21900151" />
+        <register type="NOA" address="0x00009888" value="0x23900152" />
+        <register type="NOA" address="0x00009888" value="0x25900153" />
+        <register type="NOA" address="0x00009888" value="0x27900154" />
+        <register type="NOA" address="0x00009888" value="0x29900155" />
+        <register type="NOA" address="0x00009888" value="0x2B900156" />
+        <register type="NOA" address="0x00009888" value="0x2D900157" />
+        <register type="NOA" address="0x00009888" value="0x2F90015F" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F872" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="a0c0172c-ee13-403d-99ff-2bdf6936cf14"
+       chipset="SKLGT3"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiMemoryWrites"
+             description="The total number of GTI memory writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all GTI accesses to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiMscMemoryWrites"
+             description="The total number of GTI memory writes from Multisampling Color Cache (Multisampling Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_writes"
+             units="messages"
+             symbol_name="GtiMscMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiCmdStreamerMemoryWrites"
+             description="The total number of GTI memory writes from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_writes"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiL3Bank0Writes"
+             description="The total number of GTI memory writes from L3 Bank 0 (L3 Bank 0 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_writes"
+             units="messages"
+             symbol_name="GtiL3Bank0Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank1Writes"
+             description="The total number of GTI memory writes from L3 Bank 1 (L3 Bank 1 invalidations)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_writes"
+             units="messages"
+             symbol_name="GtiL3Bank1Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank2Writes"
+             description="The total number of GTI memory writes from L3 Bank 2 (L3 Bank 2 invalidations)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_writes"
+             units="messages"
+             symbol_name="GtiL3Bank2Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank3Writes"
+             description="The total number of GTI memory writes from L3 Bank 3 (L3 Bank 3 invalidations)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_writes"
+             units="messages"
+             symbol_name="GtiL3Bank3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Writes $GtiL3Bank1Writes $GtiL3Bank2Writes $GtiL3Bank3Writes UADD UADD UADD"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRccMemoryWrites"
+             description="The total number of GTI memory writes from Render Color Cache (Render Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_writes"
+             units="messages"
+             symbol_name="GtiRccMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiSoMemoryWrites"
+             description="The total number of GTI memory writes from Stream Output."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_so_memory_writes"
+             units="messages"
+             symbol_name="GtiSoMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiStcMemoryWrites"
+             description="The total number of GTI memory writes from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_writes"
+             units="messages"
+             symbol_name="GtiStcMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GtiRczMemoryWrites"
+             description="The total number of GTI memory writes from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_writes"
+             units="messages"
+             symbol_name="GtiRczMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiHizMemoryWrites"
+             description="The total number of GTI memory writes from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_writes"
+             units="messages"
+             symbol_name="GtiHizMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901000" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900160" />
+        <register type="NOA" address="0x00009888" value="0x21900161" />
+        <register type="NOA" address="0x00009888" value="0x23900162" />
+        <register type="NOA" address="0x00009888" value="0x25900163" />
+        <register type="NOA" address="0x00009888" value="0x27900164" />
+        <register type="NOA" address="0x00009888" value="0x29900165" />
+        <register type="NOA" address="0x00009888" value="0x2B900166" />
+        <register type="NOA" address="0x00009888" value="0x2D900167" />
+        <register type="NOA" address="0x00009888" value="0x2F900150" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F822" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="52435e0b-f188-42ea-8680-21a56ee20dee"
+       chipset="SKLGT3"
+       symbol_name="ComputeExtended"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Writes 0"
+             description="The subslice 0 typed writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="typed_writes0"
+             units="messages"
+             symbol_name="TypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuTypedAtomics0"
+             description="The subslice 0 EU Typed Atomics subslice 0."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_typed_atomics0"
+             units="messages"
+             symbol_name="EuTypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Atomics 0"
+             description="The subslice 0 typed atomics."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="typed_atomics0"
+             units="messages"
+             symbol_name="TypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedAtomicsPerCacheLine"
+             description="The ratio of EU typed atomics requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedAtomics0 $TypedAtomics0 FDIV"
+             underscore_name="typed_atomics_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedAtomicsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedReads0"
+             description="The subslice 0 EU Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_untyped_reads0"
+             units="messages"
+             symbol_name="EuUntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Writes 0"
+             description="The subslice 0 untyped writes (including SLM writes)."
+             data_type="uint64"
+             equation="C 1 READ"
+             underscore_name="untyped_writes0"
+             units="messages"
+             symbol_name="UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedAtomics0"
+             description="The subslice 0 EU Untyped Atomics subslice 0."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_untyped_atomics0"
+             units="messages"
+             symbol_name="EuUntypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedWrites0"
+             description="The subslice 0 EU Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="eu_untyped_writes0"
+             units="messages"
+             symbol_name="EuUntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedWrites0"
+             description="The subslice 0 EU A64 Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_a64_untyped_writes0"
+             units="messages"
+             symbol_name="EuA64UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedWritesPerCacheLine"
+             description="The ratio of EU untyped write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuUntypedWrites0 $EuA64UntypedWrites0 UADD $UntypedWrites0 FDIV"
+             underscore_name="untyped_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedWrites0"
+             description="The subslice 0 EU Typed Writes subslice 0."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="eu_typed_writes0"
+             units="messages"
+             symbol_name="EuTypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedWritesPerCacheLine"
+             description="The ratio of EU typed write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedWrites0 $TypedWrites0 FDIV"
+             underscore_name="typed_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Reads 0"
+             description="The subslice 0 typed reads."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="typed_reads0"
+             units="messages"
+             symbol_name="TypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Reads 0"
+             description="The subslice 0 untyped reads (including SLM reads)."
+             data_type="uint64"
+             equation="C 3 READ"
+             underscore_name="untyped_reads0"
+             units="messages"
+             symbol_name="UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedReads0"
+             description="The subslice 0 EU A64 Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_a64_untyped_reads0"
+             units="messages"
+             symbol_name="EuA64UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedReads0"
+             description="The subslice 0 EU Typed Reads subslice 0."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="eu_typed_reads0"
+             units="messages"
+             symbol_name="EuTypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedReadsPerCacheLine"
+             description="The ratio of EU untyped read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuUntypedReads0 $EuA64UntypedReads0 UADD $UntypedReads0 FDIV"
+             underscore_name="untyped_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="The ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuTypedReads0 $TypedReads0 FDIV"
+             underscore_name="typed_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x141C8160" />
+        <register type="NOA" address="0x00009888" value="0x161C8015" />
+        <register type="NOA" address="0x00009888" value="0x181C0120" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4EAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B01" />
+        <register type="NOA" address="0x00009888" value="0x006C0200" />
+        <register type="NOA" address="0x00009888" value="0x026C000C" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x001C0041" />
+        <register type="NOA" address="0x00009888" value="0x061C4200" />
+        <register type="NOA" address="0x00009888" value="0x081C4443" />
+        <register type="NOA" address="0x00009888" value="0x0A1C4645" />
+        <register type="NOA" address="0x00009888" value="0x0C1C7647" />
+        <register type="NOA" address="0x00009888" value="0x041C7357" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x101C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0000" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4CAA2A" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02AA" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5515" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x11907FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900040" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900802" />
+        <register type="NOA" address="0x00009888" value="0x47900842" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900842" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900800" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FC2A" />
+        <register type="OA" address="0x00002774" value="0x0000BF00" />
+        <register type="OA" address="0x00002778" value="0x0007FC6A" />
+        <register type="OA" address="0x0000277C" value="0x0000BF00" />
+        <register type="OA" address="0x00002780" value="0x0007FC92" />
+        <register type="OA" address="0x00002784" value="0x0000BF00" />
+        <register type="OA" address="0x00002788" value="0x0007FCA2" />
+        <register type="OA" address="0x0000278C" value="0x0000BF00" />
+        <register type="OA" address="0x00002790" value="0x0007FC32" />
+        <register type="OA" address="0x00002794" value="0x0000BF00" />
+        <register type="OA" address="0x00002798" value="0x0007FC9A" />
+        <register type="OA" address="0x0000279C" value="0x0000BF00" />
+        <register type="OA" address="0x000027A0" value="0x0007FE6A" />
+        <register type="OA" address="0x000027A4" value="0x0000BF00" />
+        <register type="OA" address="0x000027A8" value="0x0007FE7A" />
+        <register type="OA" address="0x000027AC" value="0x0000BF00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics L3 Cache Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_l3_cache"
+       hw_config_guid="27076eeb-49f3-4fed-8423-c66506005c63"
+       chipset="SKLGT3"
+       symbol_name="ComputeL3Cache"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 03 Accesses"
+             description="The total number of accesses to L3 Bank 03."
+             data_type="uint64"
+             equation="B 3 READ 2 UMUL"
+             underscore_name="l3_bank03_accesses"
+             units="messages"
+             symbol_name="L3Bank03Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Accesses"
+             description="The total number of L3 accesses from all entities."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ B 2 READ B 3 READ UADD UADD UADD 2 UMUL"
+             underscore_name="l3_accesses"
+             units="messages"
+             symbol_name="L3Accesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 29 READ 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="EU FPU0 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu0_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ C 5 READ UADD"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="L3 Bank 00 Accesses"
+             description="The total number of accesses to L3 Bank 00."
+             data_type="uint64"
+             equation="C 0 READ 2 UMUL"
+             underscore_name="l3_bank00_accesses"
+             units="messages"
+             symbol_name="L3Bank00Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU0 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 19 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu0_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 14 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu1_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 00 IC Accesses"
+             description="The total number of accesses to L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 0 READ B 1 READ UADD 2 UMUL $L3Bank00Accesses UMIN"
+             underscore_name="l3_bank00_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank00IcAccesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 00 IC Hits"
+             description="The total number of hits in L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 1 READ 2 UMUL $L3Bank00IcAccesses UMIN"
+             underscore_name="l3_bank00_ic_hits"
+             units="messages"
+             symbol_name="L3Bank00IcHits"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="L3 Bank 01 Accesses"
+             description="The total number of accesses to L3 Bank 01."
+             data_type="uint64"
+             equation="C 1 READ 2 UMUL"
+             underscore_name="l3_bank01_accesses"
+             units="messages"
+             symbol_name="L3Bank01Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 20 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu1_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="L3 Bank 02 Accesses"
+             description="The total number of accesses to L3 Bank 02."
+             data_type="uint64"
+             equation="B 2 READ 2 UMUL"
+             underscore_name="l3_bank02_accesses"
+             units="messages"
+             symbol_name="L3Bank02Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="L3 Total Throughput"
+             description="The total number of GPU memory bytes transferred via L3."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Accesses 64 UMUL"
+             underscore_name="l3_total_throughput"
+             units="bytes"
+             symbol_name="L3TotalThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="EU FPU1 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C0760" />
+        <register type="NOA" address="0x00009888" value="0x1593001E" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E8020" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x006C0051" />
+        <register type="NOA" address="0x00009888" value="0x066C5000" />
+        <register type="NOA" address="0x00009888" value="0x086C5C5D" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5E5F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x061B8000" />
+        <register type="NOA" address="0x00009888" value="0x081BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1CE000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2A00" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0280" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F1500" />
+        <register type="NOA" address="0x00009888" value="0x100F0140" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C0A00" />
+        <register type="NOA" address="0x00009888" value="0x182C00A0" />
+        <register type="NOA" address="0x00009888" value="0x03933300" />
+        <register type="NOA" address="0x00009888" value="0x05930032" />
+        <register type="NOA" address="0x00009888" value="0x11930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190030F" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900063" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x53903333" />
+        <register type="NOA" address="0x00009888" value="0x43900840" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFFA" />
+        <register type="OA" address="0x00002774" value="0x0000FEFE" />
+        <register type="OA" address="0x00002778" value="0x0007FFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000FEFD" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000FBEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000FBDF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00101100" />
+        <register type="FLEX" address="0x0000E45C" value="0x00201200" />
+        <register type="FLEX" address="0x0000E55C" value="0x00301300" />
+        <register type="FLEX" address="0x0000E65C" value="0x00401400" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="8071b409-c39a-4674-94d7-32962ecfb512"
+       chipset="SKLGT3"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ C 2 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader02_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader02AccessStalledOnL3"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F0232" />
+        <register type="NOA" address="0x00009888" value="0x124F4640" />
+        <register type="NOA" address="0x00009888" value="0x106C0232" />
+        <register type="NOA" address="0x00009888" value="0x11834400" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x004F1880" />
+        <register type="NOA" address="0x00009888" value="0x024F08BB" />
+        <register type="NOA" address="0x00009888" value="0x044F001B" />
+        <register type="NOA" address="0x00009888" value="0x046C0100" />
+        <register type="NOA" address="0x00009888" value="0x066C000B" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x041B8000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025BC000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x165C8000" />
+        <register type="NOA" address="0x00009888" value="0x185C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00A0" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x062CC000" />
+        <register type="NOA" address="0x00009888" value="0x082CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x0F8305C0" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x07830000" />
+        <register type="NOA" address="0x00009888" value="0x1D950080" />
+        <register type="NOA" address="0x00009888" value="0x13928000" />
+        <register type="NOA" address="0x00009888" value="0x0F988000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x59900005" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="5e0b391e-9ea8-4901-b2ff-b64ff616c7ed"
+       chipset="SKLGT3"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C7B40" />
+        <register type="NOA" address="0x00009888" value="0x166C0020" />
+        <register type="NOA" address="0x00009888" value="0x0A603444" />
+        <register type="NOA" address="0x00009888" value="0x0A613400" />
+        <register type="NOA" address="0x00009888" value="0x1A4EA800" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C5327" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5425" />
+        <register type="NOA" address="0x00009888" value="0x006C2A00" />
+        <register type="NOA" address="0x00009888" value="0x026C285B" />
+        <register type="NOA" address="0x00009888" value="0x046C005C" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0800" />
+        <register type="NOA" address="0x00009888" value="0x0C1BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C003C" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x10600000" />
+        <register type="NOA" address="0x00009888" value="0x04600000" />
+        <register type="NOA" address="0x00009888" value="0x0C610044" />
+        <register type="NOA" address="0x00009888" value="0x10610000" />
+        <register type="NOA" address="0x00009888" value="0x06610000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02A8" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0154" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190FFC0" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900420" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900021" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900400" />
+        <register type="NOA" address="0x00009888" value="0x43900421" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_2"
+       hw_config_guid="25dc828e-1d2d-426e-9546-a1d4233cdf16"
+       chipset="SKLGT3"
+       symbol_name="L3_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Stalled"
+             description="The percentage of time in which slice0 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_stalled"
+             units="percent"
+             symbol_name="L30Bank2Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Active"
+             description="The percentage of time in which slice0 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_active"
+             units="percent"
+             symbol_name="L30Bank2Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C02E0" />
+        <register type="NOA" address="0x00009888" value="0x146C0001" />
+        <register type="NOA" address="0x00009888" value="0x0A623400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x026C3324" />
+        <register type="NOA" address="0x00009888" value="0x046C3422" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x06614000" />
+        <register type="NOA" address="0x00009888" value="0x0C620044" />
+        <register type="NOA" address="0x00009888" value="0x10620000" />
+        <register type="NOA" address="0x00009888" value="0x06620000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_3"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_3"
+       hw_config_guid="3dba9405-2d7e-4d70-8199-e734e82fd6bf"
+       chipset="SKLGT3"
+       symbol_name="L3_3"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 L3 Bank3 Stalled"
+             description="The percentage of time in which slice0 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_stalled"
+             units="percent"
+             symbol_name="L30Bank3Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank3 Active"
+             description="The percentage of time in which slice0 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_active"
+             units="percent"
+             symbol_name="L30Bank3Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C4E80" />
+        <register type="NOA" address="0x00009888" value="0x146C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A633400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x026C3321" />
+        <register type="NOA" address="0x00009888" value="0x046C342F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C2000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x06604000" />
+        <register type="NOA" address="0x00009888" value="0x0C630044" />
+        <register type="NOA" address="0x00009888" value="0x10630000" />
+        <register type="NOA" address="0x00009888" value="0x06630000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00AA" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900002" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="76935d7b-09c9-46bf-87f1-c18b4a86ebe5"
+       chipset="SKLGT3"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x102F3800" />
+        <register type="NOA" address="0x00009888" value="0x144D0500" />
+        <register type="NOA" address="0x00009888" value="0x120D03C0" />
+        <register type="NOA" address="0x00009888" value="0x140D03CF" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0004" />
+        <register type="NOA" address="0x00009888" value="0x0C4E4000" />
+        <register type="NOA" address="0x00009888" value="0x042F0480" />
+        <register type="NOA" address="0x00009888" value="0x082F0000" />
+        <register type="NOA" address="0x00009888" value="0x022F0000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0090" />
+        <register type="NOA" address="0x00009888" value="0x064D0027" />
+        <register type="NOA" address="0x00009888" value="0x004D0000" />
+        <register type="NOA" address="0x00009888" value="0x000D0D40" />
+        <register type="NOA" address="0x00009888" value="0x020D803F" />
+        <register type="NOA" address="0x00009888" value="0x040D8023" />
+        <register type="NOA" address="0x00009888" value="0x100D0000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020F0010" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0050" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41901400" />
+        <register type="NOA" address="0x00009888" value="0x43901485" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900001" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000EFFF" />
+        <register type="OA" address="0x00002778" value="0x00006000" />
+        <register type="OA" address="0x0000277C" value="0x0000F3FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler"
+       hw_config_guid="1b34c0d6-4f4c-4d7b-833f-4aaf236d87a6"
+       chipset="SKLGT3"
+       symbol_name="Sampler"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x14152C00" />
+        <register type="NOA" address="0x00009888" value="0x16150005" />
+        <register type="NOA" address="0x00009888" value="0x121600A0" />
+        <register type="NOA" address="0x00009888" value="0x14352C00" />
+        <register type="NOA" address="0x00009888" value="0x16350005" />
+        <register type="NOA" address="0x00009888" value="0x123600A0" />
+        <register type="NOA" address="0x00009888" value="0x14552C00" />
+        <register type="NOA" address="0x00009888" value="0x16550005" />
+        <register type="NOA" address="0x00009888" value="0x125600A0" />
+        <register type="NOA" address="0x00009888" value="0x062F6000" />
+        <register type="NOA" address="0x00009888" value="0x022F2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0050" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0010" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0350" />
+        <register type="NOA" address="0x00009888" value="0x0C0FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F00DA" />
+        <register type="NOA" address="0x00009888" value="0x182C0028" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x022DC000" />
+        <register type="NOA" address="0x00009888" value="0x042D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C138000" />
+        <register type="NOA" address="0x00009888" value="0x0E132000" />
+        <register type="NOA" address="0x00009888" value="0x0413C000" />
+        <register type="NOA" address="0x00009888" value="0x1C140018" />
+        <register type="NOA" address="0x00009888" value="0x0C157000" />
+        <register type="NOA" address="0x00009888" value="0x0E150078" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x04162180" />
+        <register type="NOA" address="0x00009888" value="0x02160000" />
+        <register type="NOA" address="0x00009888" value="0x04174000" />
+        <register type="NOA" address="0x00009888" value="0x0233A000" />
+        <register type="NOA" address="0x00009888" value="0x04333000" />
+        <register type="NOA" address="0x00009888" value="0x14348000" />
+        <register type="NOA" address="0x00009888" value="0x16348000" />
+        <register type="NOA" address="0x00009888" value="0x02357870" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x04360043" />
+        <register type="NOA" address="0x00009888" value="0x02360000" />
+        <register type="NOA" address="0x00009888" value="0x04371000" />
+        <register type="NOA" address="0x00009888" value="0x0E538000" />
+        <register type="NOA" address="0x00009888" value="0x00538000" />
+        <register type="NOA" address="0x00009888" value="0x06533000" />
+        <register type="NOA" address="0x00009888" value="0x1C540020" />
+        <register type="NOA" address="0x00009888" value="0x12548000" />
+        <register type="NOA" address="0x00009888" value="0x0E557000" />
+        <register type="NOA" address="0x00009888" value="0x00557800" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x06560043" />
+        <register type="NOA" address="0x00009888" value="0x02560000" />
+        <register type="NOA" address="0x00009888" value="0x06571000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900060" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="b375c985-9953-455b-bda2-b03f7594e9db"
+       chipset="SKLGT3"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12120000" />
+        <register type="NOA" address="0x00009888" value="0x12320000" />
+        <register type="NOA" address="0x00009888" value="0x12520000" />
+        <register type="NOA" address="0x00009888" value="0x002F8000" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0015" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F03A0" />
+        <register type="NOA" address="0x00009888" value="0x0C0FF000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0095" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D4000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x02108000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x02118000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x02121880" />
+        <register type="NOA" address="0x00009888" value="0x041219B5" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x02134000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x0C308000" />
+        <register type="NOA" address="0x00009888" value="0x0E304000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x0C318000" />
+        <register type="NOA" address="0x00009888" value="0x0E314000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x0C321A80" />
+        <register type="NOA" address="0x00009888" value="0x0E320033" />
+        <register type="NOA" address="0x00009888" value="0x06320031" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x0C334000" />
+        <register type="NOA" address="0x00009888" value="0x0E331000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0E508000" />
+        <register type="NOA" address="0x00009888" value="0x00508000" />
+        <register type="NOA" address="0x00009888" value="0x02504000" />
+        <register type="NOA" address="0x00009888" value="0x0E518000" />
+        <register type="NOA" address="0x00009888" value="0x00518000" />
+        <register type="NOA" address="0x00009888" value="0x02514000" />
+        <register type="NOA" address="0x00009888" value="0x0E521880" />
+        <register type="NOA" address="0x00009888" value="0x00521A80" />
+        <register type="NOA" address="0x00009888" value="0x02520033" />
+        <register type="NOA" address="0x00009888" value="0x0E534000" />
+        <register type="NOA" address="0x00009888" value="0x00534000" />
+        <register type="NOA" address="0x00009888" value="0x02531000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900062" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x00007FFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x00009FFF" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000EFFF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000F3FF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FDFF" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="3e2be2bb-884a-49bb-82c5-2358e6bd5f2d"
+       chipset="SKLGT3"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12124D60" />
+        <register type="NOA" address="0x00009888" value="0x12322E60" />
+        <register type="NOA" address="0x00009888" value="0x12524D60" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0014" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0FE000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0097" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x002D8000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x04121FB7" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x00308000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x00318000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x00321B80" />
+        <register type="NOA" address="0x00009888" value="0x0632003F" />
+        <register type="NOA" address="0x00009888" value="0x00334000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0250C000" />
+        <register type="NOA" address="0x00009888" value="0x0251C000" />
+        <register type="NOA" address="0x00009888" value="0x02521FB7" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x02535000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900063" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extra Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extra"
+       hw_config_guid="2d80a648-7b5a-4e92-bbe7-3b5c76f2e221"
+       chipset="SKLGT3"
+       symbol_name="ComputeExtra"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active including Ext Math"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing including Extended Math processing"
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ  C 5 READ C 6 READ FADD C 7 READ FADD C 2 READ FADD C 3 READ FADD C 4 READ FADD 8 FMUL FADD 100 FMUL $EuCoresTotalCount FDIV $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active_adjusted"
+             units="percent"
+             symbol_name="Fpu1ActiveAdjusted"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121203E0" />
+        <register type="NOA" address="0x00009888" value="0x123203E0" />
+        <register type="NOA" address="0x00009888" value="0x125203E0" />
+        <register type="NOA" address="0x00009888" value="0x129203E0" />
+        <register type="NOA" address="0x00009888" value="0x12B203E0" />
+        <register type="NOA" address="0x00009888" value="0x12D203E0" />
+        <register type="NOA" address="0x00009888" value="0x024EC000" />
+        <register type="NOA" address="0x00009888" value="0x044EC000" />
+        <register type="NOA" address="0x00009888" value="0x064EC000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0042" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F006D" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x042D8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06114000" />
+        <register type="NOA" address="0x00009888" value="0x06120033" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x04308000" />
+        <register type="NOA" address="0x00009888" value="0x04318000" />
+        <register type="NOA" address="0x00009888" value="0x04321980" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x04334000" />
+        <register type="NOA" address="0x00009888" value="0x04504000" />
+        <register type="NOA" address="0x00009888" value="0x04514000" />
+        <register type="NOA" address="0x00009888" value="0x04520033" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x04531000" />
+        <register type="NOA" address="0x00009888" value="0x00AF8000" />
+        <register type="NOA" address="0x00009888" value="0x0ACC0001" />
+        <register type="NOA" address="0x00009888" value="0x008D8000" />
+        <register type="NOA" address="0x00009888" value="0x028DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C8FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E8F0001" />
+        <register type="NOA" address="0x00009888" value="0x06AC8000" />
+        <register type="NOA" address="0x00009888" value="0x02AD4000" />
+        <register type="NOA" address="0x00009888" value="0x02908000" />
+        <register type="NOA" address="0x00009888" value="0x02918000" />
+        <register type="NOA" address="0x00009888" value="0x02921980" />
+        <register type="NOA" address="0x00009888" value="0x00920000" />
+        <register type="NOA" address="0x00009888" value="0x02934000" />
+        <register type="NOA" address="0x00009888" value="0x02B04000" />
+        <register type="NOA" address="0x00009888" value="0x02B14000" />
+        <register type="NOA" address="0x00009888" value="0x02B20033" />
+        <register type="NOA" address="0x00009888" value="0x00B20000" />
+        <register type="NOA" address="0x00009888" value="0x02B31000" />
+        <register type="NOA" address="0x00009888" value="0x00D08000" />
+        <register type="NOA" address="0x00009888" value="0x00D18000" />
+        <register type="NOA" address="0x00009888" value="0x00D21980" />
+        <register type="NOA" address="0x00009888" value="0x00D34000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900402" />
+        <register type="NOA" address="0x00009888" value="0x53901550" />
+        <register type="NOA" address="0x00009888" value="0x45900080" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+  </set>
+
+  <set name="Media Vme Pipe Gen9"
+       mdapi_supported_apis="MEDIA IO BB"
+       underscore_name="vme_pipe"
+       hw_config_guid="cfae9232-6ffc-42cc-a703-9790016925f0"
+       chipset="SKLGT3"
+       symbol_name="VMEPipe"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL  GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="8 A 10 READ FMUL $EuThreadsCount FDIV $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VME Busy"
+             description="The percentage of time in which VME (IME or CRE) was actively processing data."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ B 3 READ FADD 2 FDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vme_busy"
+             units="percent"
+             symbol_name="VMEBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Overview System Batch Tier2"
+             mdapi_group="VME Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A5800" />
+        <register type="NOA" address="0x00009888" value="0x161A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12180240" />
+        <register type="NOA" address="0x00009888" value="0x14180002" />
+        <register type="NOA" address="0x00009888" value="0x149A5800" />
+        <register type="NOA" address="0x00009888" value="0x169A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12980240" />
+        <register type="NOA" address="0x00009888" value="0x14980002" />
+        <register type="NOA" address="0x00009888" value="0x1A4E3FC0" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x022F8000" />
+        <register type="NOA" address="0x00009888" value="0x042F3000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C9500" />
+        <register type="NOA" address="0x00009888" value="0x0C4C002A" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0015" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C000A" />
+        <register type="NOA" address="0x00009888" value="0x04193000" />
+        <register type="NOA" address="0x00009888" value="0x081A28C1" />
+        <register type="NOA" address="0x00009888" value="0x001A0000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x0613C000" />
+        <register type="NOA" address="0x00009888" value="0x0813F000" />
+        <register type="NOA" address="0x00009888" value="0x00172000" />
+        <register type="NOA" address="0x00009888" value="0x06178000" />
+        <register type="NOA" address="0x00009888" value="0x0817A000" />
+        <register type="NOA" address="0x00009888" value="0x00180037" />
+        <register type="NOA" address="0x00009888" value="0x06180940" />
+        <register type="NOA" address="0x00009888" value="0x08180000" />
+        <register type="NOA" address="0x00009888" value="0x02180000" />
+        <register type="NOA" address="0x00009888" value="0x04183000" />
+        <register type="NOA" address="0x00009888" value="0x04AFC000" />
+        <register type="NOA" address="0x00009888" value="0x06AF3000" />
+        <register type="NOA" address="0x00009888" value="0x0ACC4000" />
+        <register type="NOA" address="0x00009888" value="0x0CCC0015" />
+        <register type="NOA" address="0x00009888" value="0x0A8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E8F4000" />
+        <register type="NOA" address="0x00009888" value="0x108F0015" />
+        <register type="NOA" address="0x00009888" value="0x16ACA000" />
+        <register type="NOA" address="0x00009888" value="0x18AC000A" />
+        <register type="NOA" address="0x00009888" value="0x06993000" />
+        <register type="NOA" address="0x00009888" value="0x0C9A28C1" />
+        <register type="NOA" address="0x00009888" value="0x009A0000" />
+        <register type="NOA" address="0x00009888" value="0x0A93F000" />
+        <register type="NOA" address="0x00009888" value="0x0C93F000" />
+        <register type="NOA" address="0x00009888" value="0x0A97A000" />
+        <register type="NOA" address="0x00009888" value="0x0C97A000" />
+        <register type="NOA" address="0x00009888" value="0x0A980977" />
+        <register type="NOA" address="0x00009888" value="0x08980000" />
+        <register type="NOA" address="0x00009888" value="0x04980000" />
+        <register type="NOA" address="0x00009888" value="0x06983000" />
+        <register type="NOA" address="0x00009888" value="0x119000FF" />
+        <register type="NOA" address="0x00009888" value="0x51900050" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x55900115" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x47900884" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900002" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00100030" />
+        <register type="OA" address="0x00002774" value="0x0000FFF9" />
+        <register type="OA" address="0x00002778" value="0x00000002" />
+        <register type="OA" address="0x0000277C" value="0x0000FFFC" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFF3" />
+        <register type="OA" address="0x00002788" value="0x00100180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FF3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00008003" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="2b985803-d3c9-4629-8a4f-634bfecba0e8"
+       chipset="SKLGT3"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810000" />
+        <register type="NOA" address="0x00009888" value="0x07810013" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930040" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
diff --git a/src/mesa/drivers/dri/i965/brw_oa_sklgt4.xml b/src/mesa/drivers/dri/i965/brw_oa_sklgt4.xml
new file mode 100644
index 0000000..10833f1
--- /dev/null
+++ b/src/mesa/drivers/dri/i965/brw_oa_sklgt4.xml
@@ -0,0 +1,10522 @@
+<?xml version="1.0"?>
+<metrics version="1491577975" merge_md5="">
+  <set name="Render Metrics Basic Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_basic"
+       hw_config_guid="bad77c24-cc64-480d-99bf-e7b740713800"
+       chipset="SKLGT4"
+       symbol_name="RenderBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Misses 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Cache Misses"
+             description="The total number of sampler cache misses in all LODs in all sampler units."
+             data_type="uint64"
+             equation="B 4 READ B 5 READ UADD B 3 READ UADD 8 UMUL"
+             underscore_name="sampler_l1_misses"
+             units="messages"
+             symbol_name="SamplerL1Misses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$SamplerL1Misses 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Sampler 0 Busy"
+             description="The percentage of time in which Sampler 0 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler0_busy"
+             units="percent"
+             symbol_name="Sampler0Busy"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler 1 Busy"
+             description="The percentage of time in which Sampler 1 has been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler1_busy"
+             units="percent"
+             symbol_name="Sampler1Busy"
+             availability="$SubsliceMask 0x12 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Samplers Busy"
+             description="The percentage of time in which samplers have been processing EU requests."
+             data_type="float"
+             max_equation="100"
+             equation="$Sampler0Busy $Sampler1Busy FMAX"
+             underscore_name="samplers_busy"
+             units="percent"
+             symbol_name="SamplersBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI Fixed Pipe Throughput"
+             description="The total number of GPU memory bytes transferred between 3D Pipeline (Command Dispatch, Input Assembly and Stream Output) and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="B 6 READ B 7 READ UADD 64 UMUL"
+             underscore_name="gti_vf_throughput"
+             units="bytes"
+             symbol_name="GtiVfThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/3D Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler 0 Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which Sampler 0 has been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             max_equation="100"
+             underscore_name="sampler0_bottleneck"
+             units="percent"
+             symbol_name="Sampler0Bottleneck"
+             availability="$SubsliceMask 0x09 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="L3 Lookup Accesses w/o IC"
+             description="The total number of L3 cache lookup accesses w/o IC."
+             data_type="uint64"
+             equation="$SamplerL1Misses $ShaderMemoryAccesses UADD"
+             underscore_name="l3_lookups"
+             units="messages"
+             symbol_name="L3Lookups"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Depth Throughput"
+             description="The total number of GPU memory bytes transferred between depth caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 0 READ C 1 READ UADD 64 UMUL"
+             underscore_name="gti_depth_throughput"
+             units="bytes"
+             symbol_name="GtiDepthThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Depth Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Samplers Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which samplers have been slowing down the pipe when processing EU requests."
+             data_type="float"
+             high_watermark="15"
+             equation="$Sampler0Bottleneck"
+             max_equation="100"
+             underscore_name="sampler_bottleneck"
+             units="percent"
+             symbol_name="SamplerBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Indicate System Frame Batch Draw"
+             mdapi_group="Sampler"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="GTI HDC TLB Lookup Throughput"
+             description="The total number of GPU memory bytes transferred between GTI and HDC, when HDC is doing TLB lookups."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_hdc_lookups_throughput"
+             units="bytes"
+             symbol_name="GtiHdcLookupsThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI RCC Throughput"
+             description="The total number of GPU memory bytes transferred between render color caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 2 READ C 3 READ UADD 64 UMUL"
+             underscore_name="gti_rcc_throughput"
+             units="bytes"
+             symbol_name="GtiRccThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/Color Cache"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C01E0" />
+        <register type="NOA" address="0x00009888" value="0x12170280" />
+        <register type="NOA" address="0x00009888" value="0x12370280" />
+        <register type="NOA" address="0x00009888" value="0x16EC01E0" />
+        <register type="NOA" address="0x00009888" value="0x176C01E0" />
+        <register type="NOA" address="0x00009888" value="0x11930317" />
+        <register type="NOA" address="0x00009888" value="0x159303DF" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x1A4E03B0" />
+        <register type="NOA" address="0x00009888" value="0x0A6C0053" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A1B4000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0001" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x042F1000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4CA400" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0002" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5600" />
+        <register type="NOA" address="0x00009888" value="0x100F0001" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x062D8000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x08133000" />
+        <register type="NOA" address="0x00009888" value="0x00170020" />
+        <register type="NOA" address="0x00009888" value="0x08170021" />
+        <register type="NOA" address="0x00009888" value="0x10170000" />
+        <register type="NOA" address="0x00009888" value="0x0633C000" />
+        <register type="NOA" address="0x00009888" value="0x06370800" />
+        <register type="NOA" address="0x00009888" value="0x10370000" />
+        <register type="NOA" address="0x00009888" value="0x1ACE0230" />
+        <register type="NOA" address="0x00009888" value="0x0AEC5300" />
+        <register type="NOA" address="0x00009888" value="0x10EC0000" />
+        <register type="NOA" address="0x00009888" value="0x1CEC0000" />
+        <register type="NOA" address="0x00009888" value="0x0A9B8000" />
+        <register type="NOA" address="0x00009888" value="0x1C9C0002" />
+        <register type="NOA" address="0x00009888" value="0x0ACC2000" />
+        <register type="NOA" address="0x00009888" value="0x0CCC0002" />
+        <register type="NOA" address="0x00009888" value="0x088D8000" />
+        <register type="NOA" address="0x00009888" value="0x0A8D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E8F1000" />
+        <register type="NOA" address="0x00009888" value="0x108F0001" />
+        <register type="NOA" address="0x00009888" value="0x16AC8800" />
+        <register type="NOA" address="0x00009888" value="0x1B4E0020" />
+        <register type="NOA" address="0x00009888" value="0x096C5300" />
+        <register type="NOA" address="0x00009888" value="0x116C0000" />
+        <register type="NOA" address="0x00009888" value="0x1D6C0000" />
+        <register type="NOA" address="0x00009888" value="0x091B8000" />
+        <register type="NOA" address="0x00009888" value="0x1B1C8000" />
+        <register type="NOA" address="0x00009888" value="0x0B4C2000" />
+        <register type="NOA" address="0x00009888" value="0x090D8000" />
+        <register type="NOA" address="0x00009888" value="0x0F0F1000" />
+        <register type="NOA" address="0x00009888" value="0x172C0800" />
+        <register type="NOA" address="0x00009888" value="0x0D933031" />
+        <register type="NOA" address="0x00009888" value="0x0F933E3F" />
+        <register type="NOA" address="0x00009888" value="0x01933D00" />
+        <register type="NOA" address="0x00009888" value="0x0393073C" />
+        <register type="NOA" address="0x00009888" value="0x0593000E" />
+        <register type="NOA" address="0x00009888" value="0x1D930000" />
+        <register type="NOA" address="0x00009888" value="0x19930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x2B908000" />
+        <register type="NOA" address="0x00009888" value="0x2D908000" />
+        <register type="NOA" address="0x00009888" value="0x2F908000" />
+        <register type="NOA" address="0x00009888" value="0x31908000" />
+        <register type="NOA" address="0x00009888" value="0x15908000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190003F" />
+        <register type="NOA" address="0x00009888" value="0x5190FF30" />
+        <register type="NOA" address="0x00009888" value="0x41900060" />
+        <register type="NOA" address="0x00009888" value="0x55903033" />
+        <register type="NOA" address="0x00009888" value="0x45901421" />
+        <register type="NOA" address="0x00009888" value="0x47900803" />
+        <register type="NOA" address="0x00009888" value="0x5790FFF1" />
+        <register type="NOA" address="0x00009888" value="0x49900001" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x5990000F" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x5390FFFF" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Basic Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_basic"
+       hw_config_guid="7277228f-e7f3-4743-945a-6a2049d11377"
+       chipset="SKLGT4"
+       symbol_name="ComputeBasic"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 6 READ B 7 READ C 0 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_read"
+             units="bytes"
+             symbol_name="UntypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Bytes Written"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 3 READ B 4 READ B 5 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_written"
+             units="bytes"
+             symbol_name="TypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Untyped Writes"
+             description="The total number of untyped memory bytes written via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 1 READ C 2 READ C 3 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="untyped_bytes_written"
+             units="bytes"
+             symbol_name="UntypedBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="Typed Bytes Read"
+             description="The total number of typed memory bytes read via Data Port."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="B 0 READ B 1 READ B 2 READ UADD UADD $EuSlicesTotalCount 64 UMUL UMUL"
+             underscore_name="typed_bytes_read"
+             units="bytes"
+             symbol_name="TypedBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 5 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO OCL BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F00E0" />
+        <register type="NOA" address="0x00009888" value="0x124F1C00" />
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E0820" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x064F0900" />
+        <register type="NOA" address="0x00009888" value="0x084F0032" />
+        <register type="NOA" address="0x00009888" value="0x0A4F1891" />
+        <register type="NOA" address="0x00009888" value="0x0C4F0E00" />
+        <register type="NOA" address="0x00009888" value="0x0E4F003C" />
+        <register type="NOA" address="0x00009888" value="0x004F0D80" />
+        <register type="NOA" address="0x00009888" value="0x024F003B" />
+        <register type="NOA" address="0x00009888" value="0x006C0002" />
+        <register type="NOA" address="0x00009888" value="0x086C0100" />
+        <register type="NOA" address="0x00009888" value="0x0C6C000C" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B00" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x081B8000" />
+        <register type="NOA" address="0x00009888" value="0x0C1B4000" />
+        <register type="NOA" address="0x00009888" value="0x0E1B8000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C8000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0024" />
+        <register type="NOA" address="0x00009888" value="0x065B8000" />
+        <register type="NOA" address="0x00009888" value="0x085B4000" />
+        <register type="NOA" address="0x00009888" value="0x0A5BC000" />
+        <register type="NOA" address="0x00009888" value="0x0C5B8000" />
+        <register type="NOA" address="0x00009888" value="0x0E5B4000" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C6000" />
+        <register type="NOA" address="0x00009888" value="0x1C5C001B" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0208" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2CC000" />
+        <register type="NOA" address="0x00009888" value="0x162CFB00" />
+        <register type="NOA" address="0x00009888" value="0x182C00BE" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x19900157" />
+        <register type="NOA" address="0x00009888" value="0x1B900158" />
+        <register type="NOA" address="0x00009888" value="0x1D900105" />
+        <register type="NOA" address="0x00009888" value="0x1F900103" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x11900FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900821" />
+        <register type="NOA" address="0x00009888" value="0x47900802" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900802" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900002" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900422" />
+        <register type="NOA" address="0x00009888" value="0x53905555" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Render Metrics for 3D Pipeline Profile Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="render_pipe_profile"
+       hw_config_guid="463c668c-3f60-49b6-8f85-d995b635b3b2"
+       chipset="SKLGT4"
+       symbol_name="RenderPipeProfile"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which vertex shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_bottleneck"
+             units="percent"
+             symbol_name="VsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Hi-Depth Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which early hierarchical depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hi_depth_bottleneck"
+             units="percent"
+             symbol_name="HiDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which geometry shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gs_bottleneck"
+             units="percent"
+             symbol_name="GsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Geometry Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="BC Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which barycentric coordinates calculation pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="bc_bottleneck"
+             units="percent"
+             symbol_name="BcBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Barycentric Calc"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Stall"
+             description="The percentage of time in which hull stall pipeline stage was stalled."
+             data_type="float"
+             equation="C 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_stall"
+             units="percent"
+             symbol_name="HsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Hull Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="VF Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which vertex fetch pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="C 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vf_bottleneck"
+             units="percent"
+             symbol_name="VfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Input Assembler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Strip-Fans Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which strip-fans pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="10"
+             equation="B 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_bottleneck"
+             units="percent"
+             symbol_name="SfBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SF Stall"
+             description="The percentage of time in which strip-fans pipeline stage was stalled."
+             data_type="float"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sf_stall"
+             units="percent"
+             symbol_name="SfStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Strip-Fans"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="HS Bottleneck"
+             low_watermark="3"
+             description="The percentage of time in which hull shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="9"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="hs_bottleneck"
+             units="percent"
+             symbol_name="HsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Hull Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CL Stall"
+             description="The percentage of time in which clipper pipeline stage was stalled."
+             data_type="float"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_stall"
+             units="percent"
+             symbol_name="ClStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Clipper"
+             />
+    <counter name="SO Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which stream output pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_bottleneck"
+             units="percent"
+             symbol_name="SoBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Stream Output"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="DS Bottleneck"
+             low_watermark="5"
+             description="The percentage of time in which domain shader pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="15"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_bottleneck"
+             units="percent"
+             symbol_name="DsBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Domain Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Clipper Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which clipper pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="cl_bottleneck"
+             units="percent"
+             symbol_name="ClBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Clipper"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Stall"
+             description="The percentage of time in which domain shader pipeline stage was stalled."
+             data_type="float"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ds_stall"
+             units="percent"
+             symbol_name="DsStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Domain Shader"
+             />
+    <counter name="Early Depth Bottleneck"
+             low_watermark="10"
+             description="The percentage of time in which early depth test pipeline stage was slowing down the 3D pipeline."
+             data_type="float"
+             high_watermark="30"
+             equation="B 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="early_depth_bottleneck"
+             units="percent"
+             symbol_name="EarlyDepthBottleneck"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Indicate Draw"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SO Stall"
+             description="The percentage of time in which stream-output pipeline stage was stalled."
+             data_type="float"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="so_stall"
+             units="percent"
+             symbol_name="SoStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Correlate Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Stream Output"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x0C0E001F" />
+        <register type="NOA" address="0x00009888" value="0x0A0F0000" />
+        <register type="NOA" address="0x00009888" value="0x10116800" />
+        <register type="NOA" address="0x00009888" value="0x178A03E0" />
+        <register type="NOA" address="0x00009888" value="0x11824C00" />
+        <register type="NOA" address="0x00009888" value="0x11830020" />
+        <register type="NOA" address="0x00009888" value="0x13840020" />
+        <register type="NOA" address="0x00009888" value="0x11850019" />
+        <register type="NOA" address="0x00009888" value="0x11860007" />
+        <register type="NOA" address="0x00009888" value="0x01870C40" />
+        <register type="NOA" address="0x00009888" value="0x17880000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0040" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x040D4000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020E5400" />
+        <register type="NOA" address="0x00009888" value="0x000E0000" />
+        <register type="NOA" address="0x00009888" value="0x080F0040" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x100F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0040" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06110012" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x01898000" />
+        <register type="NOA" address="0x00009888" value="0x0D890100" />
+        <register type="NOA" address="0x00009888" value="0x03898000" />
+        <register type="NOA" address="0x00009888" value="0x09808000" />
+        <register type="NOA" address="0x00009888" value="0x0B808000" />
+        <register type="NOA" address="0x00009888" value="0x0380C000" />
+        <register type="NOA" address="0x00009888" value="0x0F8A0075" />
+        <register type="NOA" address="0x00009888" value="0x1D8A0000" />
+        <register type="NOA" address="0x00009888" value="0x118A8000" />
+        <register type="NOA" address="0x00009888" value="0x1B8A4000" />
+        <register type="NOA" address="0x00009888" value="0x138A8000" />
+        <register type="NOA" address="0x00009888" value="0x1D81A000" />
+        <register type="NOA" address="0x00009888" value="0x15818000" />
+        <register type="NOA" address="0x00009888" value="0x17818000" />
+        <register type="NOA" address="0x00009888" value="0x0B820030" />
+        <register type="NOA" address="0x00009888" value="0x07828000" />
+        <register type="NOA" address="0x00009888" value="0x0D824000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x05824000" />
+        <register type="NOA" address="0x00009888" value="0x0D830003" />
+        <register type="NOA" address="0x00009888" value="0x0583000C" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x03838000" />
+        <register type="NOA" address="0x00009888" value="0x07838000" />
+        <register type="NOA" address="0x00009888" value="0x0B840980" />
+        <register type="NOA" address="0x00009888" value="0x03844D80" />
+        <register type="NOA" address="0x00009888" value="0x11840000" />
+        <register type="NOA" address="0x00009888" value="0x09848000" />
+        <register type="NOA" address="0x00009888" value="0x09850080" />
+        <register type="NOA" address="0x00009888" value="0x03850003" />
+        <register type="NOA" address="0x00009888" value="0x01850000" />
+        <register type="NOA" address="0x00009888" value="0x07860000" />
+        <register type="NOA" address="0x00009888" value="0x0F860400" />
+        <register type="NOA" address="0x00009888" value="0x09870032" />
+        <register type="NOA" address="0x00009888" value="0x01888052" />
+        <register type="NOA" address="0x00009888" value="0x11880000" />
+        <register type="NOA" address="0x00009888" value="0x09884000" />
+        <register type="NOA" address="0x00009888" value="0x1B931001" />
+        <register type="NOA" address="0x00009888" value="0x1D930001" />
+        <register type="NOA" address="0x00009888" value="0x19934000" />
+        <register type="NOA" address="0x00009888" value="0x1B958000" />
+        <register type="NOA" address="0x00009888" value="0x1D950094" />
+        <register type="NOA" address="0x00009888" value="0x19958000" />
+        <register type="NOA" address="0x00009888" value="0x09E58000" />
+        <register type="NOA" address="0x00009888" value="0x0BE58000" />
+        <register type="NOA" address="0x00009888" value="0x03E5C000" />
+        <register type="NOA" address="0x00009888" value="0x0592C000" />
+        <register type="NOA" address="0x00009888" value="0x0B928000" />
+        <register type="NOA" address="0x00009888" value="0x0D924000" />
+        <register type="NOA" address="0x00009888" value="0x0F924000" />
+        <register type="NOA" address="0x00009888" value="0x11928000" />
+        <register type="NOA" address="0x00009888" value="0x1392C000" />
+        <register type="NOA" address="0x00009888" value="0x09924000" />
+        <register type="NOA" address="0x00009888" value="0x01985000" />
+        <register type="NOA" address="0x00009888" value="0x07988000" />
+        <register type="NOA" address="0x00009888" value="0x09981000" />
+        <register type="NOA" address="0x00009888" value="0x0B982000" />
+        <register type="NOA" address="0x00009888" value="0x0D982000" />
+        <register type="NOA" address="0x00009888" value="0x0F989000" />
+        <register type="NOA" address="0x00009888" value="0x05982000" />
+        <register type="NOA" address="0x00009888" value="0x13904000" />
+        <register type="NOA" address="0x00009888" value="0x21904000" />
+        <register type="NOA" address="0x00009888" value="0x23904000" />
+        <register type="NOA" address="0x00009888" value="0x25908000" />
+        <register type="NOA" address="0x00009888" value="0x27904000" />
+        <register type="NOA" address="0x00009888" value="0x29908000" />
+        <register type="NOA" address="0x00009888" value="0x2B904000" />
+        <register type="NOA" address="0x00009888" value="0x2F904000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x15904000" />
+        <register type="NOA" address="0x00009888" value="0x17908000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B904000" />
+        <register type="NOA" address="0x00009888" value="0x1190C080" />
+        <register type="NOA" address="0x00009888" value="0x51901110" />
+        <register type="NOA" address="0x00009888" value="0x41900440" />
+        <register type="NOA" address="0x00009888" value="0x55901111" />
+        <register type="NOA" address="0x00009888" value="0x45900400" />
+        <register type="NOA" address="0x00009888" value="0x47900C21" />
+        <register type="NOA" address="0x00009888" value="0x57901411" />
+        <register type="NOA" address="0x00009888" value="0x49900042" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900024" />
+        <register type="NOA" address="0x00009888" value="0x59900001" />
+        <register type="NOA" address="0x00009888" value="0x43900841" />
+        <register type="NOA" address="0x00009888" value="0x53900411" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFEA" />
+        <register type="OA" address="0x00002774" value="0x00007FFC" />
+        <register type="OA" address="0x00002778" value="0x0007AFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000F5FD" />
+        <register type="OA" address="0x00002780" value="0x00079FFA" />
+        <register type="OA" address="0x00002784" value="0x0000F3FB" />
+        <register type="OA" address="0x00002788" value="0x0007BF7A" />
+        <register type="OA" address="0x0000278C" value="0x0000F7E7" />
+        <register type="OA" address="0x00002790" value="0x0007FEFA" />
+        <register type="OA" address="0x00002794" value="0x0000F7CF" />
+        <register type="OA" address="0x00002798" value="0x00077FFA" />
+        <register type="OA" address="0x0000279C" value="0x0000EFDF" />
+        <register type="OA" address="0x000027A0" value="0x0006FFFA" />
+        <register type="OA" address="0x000027A4" value="0x0000CFBF" />
+        <register type="OA" address="0x000027A8" value="0x0003FFFA" />
+        <register type="OA" address="0x000027AC" value="0x00005F7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Reads Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_reads"
+       hw_config_guid="3ae6e74c-72c3-4040-9bd0-7961430b8cc8"
+       chipset="SKLGT4"
+       symbol_name="MemoryReads"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank0Reads"
+             description="The total number of GTI memory reads from L3 Bank 0 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_reads"
+             units="messages"
+             symbol_name="GtiL3Bank0Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all accesses from GTI to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiL3Bank3Reads"
+             description="The total number of GTI memory reads from L3 Bank 3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_reads"
+             units="messages"
+             symbol_name="GtiL3Bank3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiRsMemoryReads"
+             description="The total number of GTI memory reads from Resource Streamer."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_rs_memory_reads"
+             units="messages"
+             symbol_name="GtiRsMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Resource Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiHizMemoryReads"
+             description="The total number of GTI memory reads from Hierarchical Depth Cache (Hi-Depth Cache misses)."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_reads"
+             units="messages"
+             symbol_name="GtiHizMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="GtiRccMemoryReads"
+             description="The total number of GTI memory reads from Render Color Cache (Render Color Cache misses)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_reads"
+             units="messages"
+             symbol_name="GtiRccMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Bank1Reads"
+             description="The total number of GTI memory reads from L3 Bank 1 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_reads"
+             units="messages"
+             symbol_name="GtiL3Bank1Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiCmdStreamerMemoryReads"
+             description="The total number of GTI memory reads from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_reads"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="GtiL3Bank2Reads"
+             description="The total number of GTI memory reads from L3 Bank 2 (L3 Cache misses)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_reads"
+             units="messages"
+             symbol_name="GtiL3Bank2Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiMemoryReads"
+             description="The total number of GTI memory reads."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_reads"
+             units="messages"
+             symbol_name="GtiMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="GtiRczMemoryReads"
+             description="The total number of GTI memory reads from Render Depth Cache (Render Depth Cache misses)."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_reads"
+             units="messages"
+             symbol_name="GtiRczMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="GtiMscMemoryReads"
+             description="The total number of GTI memory reads from Multisampling Color Cache (Multisampling Color Cache misses)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_reads"
+             units="messages"
+             symbol_name="GtiMscMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiVfMemoryReads"
+             description="The total number of GTI memory reads from Vertex Fetch."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="gti_vf_memory_reads"
+             units="messages"
+             symbol_name="GtiVfMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Vertex Fetch"
+             />
+    <counter name="GtiStcMemoryReads"
+             description="The total number of GTI memory reads from Stencil Cache (Stencil Cache misses)."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_reads"
+             units="messages"
+             symbol_name="GtiStcMemoryReads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiL3Reads"
+             description="The total number of GTI memory reads from L3 (L3 Cache misses)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Reads $GtiL3Bank1Reads $GtiL3Bank2Reads $GtiL3Bank3Reads UADD UADD UADD"
+             underscore_name="gti_l3_reads"
+             units="messages"
+             symbol_name="GtiL3Reads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F900064" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900150" />
+        <register type="NOA" address="0x00009888" value="0x21900151" />
+        <register type="NOA" address="0x00009888" value="0x23900152" />
+        <register type="NOA" address="0x00009888" value="0x25900153" />
+        <register type="NOA" address="0x00009888" value="0x27900154" />
+        <register type="NOA" address="0x00009888" value="0x29900155" />
+        <register type="NOA" address="0x00009888" value="0x2B900156" />
+        <register type="NOA" address="0x00009888" value="0x2D900157" />
+        <register type="NOA" address="0x00009888" value="0x2F90015F" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F872" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Memory Writes Distribution Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="memory_writes"
+       hw_config_guid="055f256d-4052-467c-8dec-6064a4806433"
+       chipset="SKLGT4"
+       symbol_name="MemoryWrites"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiMemoryWrites"
+             description="The total number of GTI memory writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="gti_memory_writes"
+             units="messages"
+             symbol_name="GtiMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="GtiRingAccesses"
+             description="The total number of all GTI accesses to the ring."
+             data_type="uint64"
+             equation="C 3 READ 2 UMUL"
+             underscore_name="gti_ring_accesses"
+             units="messages"
+             symbol_name="GtiRingAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="GtiMscMemoryWrites"
+             description="The total number of GTI memory writes from Multisampling Color Cache (Multisampling Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="gti_msc_memory_writes"
+             units="messages"
+             symbol_name="GtiMscMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="GtiCmdStreamerMemoryWrites"
+             description="The total number of GTI memory writes from Command Streamer."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="gti_cmd_streamer_memory_writes"
+             units="messages"
+             symbol_name="GtiCmdStreamerMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Command Streamer"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GtiL3Bank0Writes"
+             description="The total number of GTI memory writes from L3 Bank 0 (L3 Bank 0 invalidations)."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="gti_l3_bank0_writes"
+             units="messages"
+             symbol_name="GtiL3Bank0Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank1Writes"
+             description="The total number of GTI memory writes from L3 Bank 1 (L3 Bank 1 invalidations)."
+             data_type="uint64"
+             equation="C 5 READ"
+             underscore_name="gti_l3_bank1_writes"
+             units="messages"
+             symbol_name="GtiL3Bank1Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank2Writes"
+             description="The total number of GTI memory writes from L3 Bank 2 (L3 Bank 2 invalidations)."
+             data_type="uint64"
+             equation="C 6 READ"
+             underscore_name="gti_l3_bank2_writes"
+             units="messages"
+             symbol_name="GtiL3Bank2Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Bank3Writes"
+             description="The total number of GTI memory writes from L3 Bank 3 (L3 Bank 3 invalidations)."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="gti_l3_bank3_writes"
+             units="messages"
+             symbol_name="GtiL3Bank3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GtiL3Writes"
+             description="The total number of GTI memory writes from L3 (L3 invalidations)."
+             data_type="uint64"
+             equation="$GtiL3Bank0Writes $GtiL3Bank1Writes $GtiL3Bank2Writes $GtiL3Bank3Writes UADD UADD UADD"
+             underscore_name="gti_l3_writes"
+             units="messages"
+             symbol_name="GtiL3Writes"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/L3"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiRccMemoryWrites"
+             description="The total number of GTI memory writes from Render Color Cache (Render Color Cache invalidations)."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="gti_rcc_memory_writes"
+             units="messages"
+             symbol_name="GtiRccMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Color Cache"
+             />
+    <counter name="GtiSoMemoryWrites"
+             description="The total number of GTI memory writes from Stream Output."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="gti_so_memory_writes"
+             units="messages"
+             symbol_name="GtiSoMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/3D Pipe/Stream Output"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GtiStcMemoryWrites"
+             description="The total number of GTI memory writes from Stencil Cache."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="gti_stc_memory_writes"
+             units="messages"
+             symbol_name="GtiStcMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="GtiRczMemoryWrites"
+             description="The total number of GTI memory writes from Render Depth Cache."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="gti_rcz_memory_writes"
+             units="messages"
+             symbol_name="GtiRczMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="GtiHizMemoryWrites"
+             description="The total number of GTI memory writes from Hierarchical Depth Cache."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="gti_hiz_memory_writes"
+             units="messages"
+             symbol_name="GtiHizMemoryWrites"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GTI/Depth Cache"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810C00" />
+        <register type="NOA" address="0x00009888" value="0x1381001A" />
+        <register type="NOA" address="0x00009888" value="0x37906800" />
+        <register type="NOA" address="0x00009888" value="0x3F901000" />
+        <register type="NOA" address="0x00009888" value="0x03811300" />
+        <register type="NOA" address="0x00009888" value="0x05811B12" />
+        <register type="NOA" address="0x00009888" value="0x0781001A" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x17810000" />
+        <register type="NOA" address="0x00009888" value="0x19810000" />
+        <register type="NOA" address="0x00009888" value="0x1B810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930055" />
+        <register type="NOA" address="0x00009888" value="0x03E58000" />
+        <register type="NOA" address="0x00009888" value="0x05E5C000" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x13900160" />
+        <register type="NOA" address="0x00009888" value="0x21900161" />
+        <register type="NOA" address="0x00009888" value="0x23900162" />
+        <register type="NOA" address="0x00009888" value="0x25900163" />
+        <register type="NOA" address="0x00009888" value="0x27900164" />
+        <register type="NOA" address="0x00009888" value="0x29900165" />
+        <register type="NOA" address="0x00009888" value="0x2B900166" />
+        <register type="NOA" address="0x00009888" value="0x2D900167" />
+        <register type="NOA" address="0x00009888" value="0x2F900150" />
+        <register type="NOA" address="0x00009888" value="0x31900105" />
+        <register type="NOA" address="0x00009888" value="0x15900103" />
+        <register type="NOA" address="0x00009888" value="0x17900101" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1D908000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C60" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900C63" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C63" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900063" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x0000272C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002728" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x0000271C" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002718" value="0xFFFFFFFF" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x0000274C" value="0x86543210" />
+        <register type="OA" address="0x00002748" value="0x86543210" />
+        <register type="OA" address="0x00002744" value="0x00006667" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x0000275C" value="0x86543210" />
+        <register type="OA" address="0x00002758" value="0x86543210" />
+        <register type="OA" address="0x00002754" value="0x00006465" />
+        <register type="OA" address="0x00002750" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007F81A" />
+        <register type="OA" address="0x00002774" value="0x0000FE00" />
+        <register type="OA" address="0x00002778" value="0x0007F82A" />
+        <register type="OA" address="0x0000277C" value="0x0000FE00" />
+        <register type="OA" address="0x00002780" value="0x0007F822" />
+        <register type="OA" address="0x00002784" value="0x0000FE00" />
+        <register type="OA" address="0x00002788" value="0x0007F8BA" />
+        <register type="OA" address="0x0000278C" value="0x0000FE00" />
+        <register type="OA" address="0x00002790" value="0x0007F87A" />
+        <register type="OA" address="0x00002794" value="0x0000FE00" />
+        <register type="OA" address="0x00002798" value="0x0007F8EA" />
+        <register type="OA" address="0x0000279C" value="0x0000FE00" />
+        <register type="OA" address="0x000027A0" value="0x0007F8E2" />
+        <register type="OA" address="0x000027A4" value="0x0000FE00" />
+        <register type="OA" address="0x000027A8" value="0x0007F8F2" />
+        <register type="OA" address="0x000027AC" value="0x0000FE00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00015014" />
+        <register type="FLEX" address="0x0000E658" value="0x00025024" />
+        <register type="FLEX" address="0x0000E758" value="0x00035034" />
+        <register type="FLEX" address="0x0000E45C" value="0x00045044" />
+        <register type="FLEX" address="0x0000E55C" value="0x00055054" />
+        <register type="FLEX" address="0x0000E65C" value="0x00065064" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extended Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extended"
+       hw_config_guid="753972d4-87cd-4460-824d-754463ac5054"
+       chipset="SKLGT4"
+       symbol_name="ComputeExtended"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Typed Writes 0"
+             description="The subslice 0 typed writes."
+             data_type="uint64"
+             equation="C 0 READ"
+             underscore_name="typed_writes0"
+             units="messages"
+             symbol_name="TypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuTypedAtomics0"
+             description="The subslice 0 EU Typed Atomics subslice 0."
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="eu_typed_atomics0"
+             units="messages"
+             symbol_name="EuTypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Atomics 0"
+             description="The subslice 0 typed atomics."
+             data_type="uint64"
+             equation="C 4 READ"
+             underscore_name="typed_atomics0"
+             units="messages"
+             symbol_name="TypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedAtomicsPerCacheLine"
+             description="The ratio of EU typed atomics requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedAtomics0 $TypedAtomics0 FDIV"
+             underscore_name="typed_atomics_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedAtomicsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedReads0"
+             description="The subslice 0 EU Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="eu_untyped_reads0"
+             units="messages"
+             symbol_name="EuUntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Writes 0"
+             description="The subslice 0 untyped writes (including SLM writes)."
+             data_type="uint64"
+             equation="C 1 READ"
+             underscore_name="untyped_writes0"
+             units="messages"
+             symbol_name="UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedAtomics0"
+             description="The subslice 0 EU Untyped Atomics subslice 0."
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="eu_untyped_atomics0"
+             units="messages"
+             symbol_name="EuUntypedAtomics0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuUntypedWrites0"
+             description="The subslice 0 EU Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="eu_untyped_writes0"
+             units="messages"
+             symbol_name="EuUntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedWrites0"
+             description="The subslice 0 EU A64 Untyped Writes subslice 0."
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="eu_a64_untyped_writes0"
+             units="messages"
+             symbol_name="EuA64UntypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedWritesPerCacheLine"
+             description="The ratio of EU untyped write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuUntypedWrites0 $EuA64UntypedWrites0 UADD $UntypedWrites0 FDIV"
+             underscore_name="untyped_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedWrites0"
+             description="The subslice 0 EU Typed Writes subslice 0."
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="eu_typed_writes0"
+             units="messages"
+             symbol_name="EuTypedWrites0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedWritesPerCacheLine"
+             description="The ratio of EU typed write requests to L3 cache line writes."
+             data_type="float"
+             equation="$EuTypedWrites0 $TypedWrites0 FDIV"
+             underscore_name="typed_writes_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedWritesPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Typed Reads 0"
+             description="The subslice 0 typed reads."
+             data_type="uint64"
+             equation="C 2 READ"
+             underscore_name="typed_reads0"
+             units="messages"
+             symbol_name="TypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Untyped Reads 0"
+             description="The subslice 0 untyped reads (including SLM reads)."
+             data_type="uint64"
+             equation="C 3 READ"
+             underscore_name="untyped_reads0"
+             units="messages"
+             symbol_name="UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EuA64UntypedReads0"
+             description="The subslice 0 EU A64 Untyped Reads subslice 0."
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="eu_a64_untyped_reads0"
+             units="messages"
+             symbol_name="EuA64UntypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ 8 UMUL $EuCoresTotalCount UDIV $EuThreadsCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EuTypedReads0"
+             description="The subslice 0 EU Typed Reads subslice 0."
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="eu_typed_reads0"
+             units="messages"
+             symbol_name="EuTypedReads0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="UntypedReadsPerCacheLine"
+             description="The ratio of EU untyped read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuUntypedReads0 $EuA64UntypedReads0 UADD $UntypedReads0 FDIV"
+             underscore_name="untyped_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="UntypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="TypedReadsPerCacheLine"
+             description="The ratio of EU typed read requests to L3 cache line reads."
+             data_type="float"
+             equation="$EuTypedReads0 $TypedReads0 FDIV"
+             underscore_name="typed_reads_per_cache_line"
+             units="eu sends to l3 cache lines"
+             symbol_name="TypedReadsPerCacheLine"
+             semantic_type="ratio"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="subslice"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x106C00E0" />
+        <register type="NOA" address="0x00009888" value="0x141C8160" />
+        <register type="NOA" address="0x00009888" value="0x161C8015" />
+        <register type="NOA" address="0x00009888" value="0x181C0120" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4EAAA0" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E6C0B01" />
+        <register type="NOA" address="0x00009888" value="0x006C0200" />
+        <register type="NOA" address="0x00009888" value="0x026C000C" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x001C0041" />
+        <register type="NOA" address="0x00009888" value="0x061C4200" />
+        <register type="NOA" address="0x00009888" value="0x081C4443" />
+        <register type="NOA" address="0x00009888" value="0x0A1C4645" />
+        <register type="NOA" address="0x00009888" value="0x0C1C7647" />
+        <register type="NOA" address="0x00009888" value="0x041C7357" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x101C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0000" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4CAA2A" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02AA" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x000DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5515" />
+        <register type="NOA" address="0x00009888" value="0x100F0155" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x11907FFF" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900040" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900802" />
+        <register type="NOA" address="0x00009888" value="0x47900842" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900842" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x43900800" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FC2A" />
+        <register type="OA" address="0x00002774" value="0x0000BF00" />
+        <register type="OA" address="0x00002778" value="0x0007FC6A" />
+        <register type="OA" address="0x0000277C" value="0x0000BF00" />
+        <register type="OA" address="0x00002780" value="0x0007FC92" />
+        <register type="OA" address="0x00002784" value="0x0000BF00" />
+        <register type="OA" address="0x00002788" value="0x0007FCA2" />
+        <register type="OA" address="0x0000278C" value="0x0000BF00" />
+        <register type="OA" address="0x00002790" value="0x0007FC32" />
+        <register type="OA" address="0x00002794" value="0x0000BF00" />
+        <register type="OA" address="0x00002798" value="0x0007FC9A" />
+        <register type="OA" address="0x0000279C" value="0x0000BF00" />
+        <register type="OA" address="0x000027A0" value="0x0007FE6A" />
+        <register type="OA" address="0x000027A4" value="0x0000BF00" />
+        <register type="OA" address="0x000027A8" value="0x0007FE7A" />
+        <register type="OA" address="0x000027AC" value="0x0000BF00" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00778008" />
+        <register type="FLEX" address="0x0000E45C" value="0x00088078" />
+        <register type="FLEX" address="0x0000E55C" value="0x00808708" />
+        <register type="FLEX" address="0x0000E65C" value="0x00A08908" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics L3 Cache Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_l3_cache"
+       hw_config_guid="4e4392e9-8f73-457b-ab44-b49f7a0c733b"
+       chipset="SKLGT4"
+       symbol_name="ComputeL3Cache"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 03 Accesses"
+             description="The total number of accesses to L3 Bank 03."
+             data_type="uint64"
+             equation="B 3 READ 2 UMUL"
+             underscore_name="l3_bank03_accesses"
+             units="messages"
+             symbol_name="L3Bank03Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="L3 Accesses"
+             description="The total number of L3 accesses from all entities."
+             data_type="uint64"
+             equation="C 0 READ C 1 READ B 2 READ B 3 READ UADD UADD UADD 2 UMUL"
+             underscore_name="l3_accesses"
+             units="messages"
+             symbol_name="L3Accesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Sampler Throughput"
+             description="The total number of GPU memory bytes transferred between samplers and L3 caches."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="A 29 READ 64 UMUL"
+             underscore_name="l3_sampler_throughput"
+             units="bytes"
+             symbol_name="L3SamplerThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Sampler"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu0_active"
+             units="percent"
+             symbol_name="Fpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU AVG IPC Rate"
+             description="The average rate of IPC calculated for 2 FPU pipelines."
+             data_type="float"
+             max_equation="2"
+             equation="A 9 READ  A 10 READ  A 11 READ UADD  A 9 READ USUB FDIV 1 FADD"
+             underscore_name="eu_avg_ipc_rate"
+             units="number"
+             symbol_name="EuAvgIpcRate"
+             semantic_type="ratio"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="EU FPU0 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 13 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu0_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Misses"
+             description="The total number of L3 misses."
+             data_type="uint64"
+             equation="C 4 READ C 5 READ UADD"
+             underscore_name="l3_misses"
+             units="messages"
+             symbol_name="L3Misses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/TAG"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="L3 Bank 00 Accesses"
+             description="The total number of accesses to L3 Bank 00."
+             data_type="uint64"
+             equation="C 0 READ 2 UMUL"
+             underscore_name="l3_bank00_accesses"
+             units="messages"
+             symbol_name="L3Bank00Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="EU FPU0 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 19 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu0_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Hybrid Instruction"
+             description="The percentage of time in which execution units were actively processing hybrid instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 14 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_hybrid_fpu1_instruction"
+             units="percent"
+             symbol_name="EuHybridFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Read Throughput"
+             description="The total number of GPU memory bytes read from GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL"
+             equation="C 6 READ 64 UMUL"
+             underscore_name="gti_read_throughput"
+             units="bytes"
+             symbol_name="GtiReadThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI L3 Throughput"
+             description="The total number of GPU memory bytes transferred between L3 caches and GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="C 4 READ C 5 READ UADD 64 UMUL"
+             underscore_name="gti_l3_throughput"
+             units="bytes"
+             symbol_name="GtiL3Throughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Bank 00 IC Accesses"
+             description="The total number of accesses to L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 0 READ B 1 READ UADD 2 UMUL $L3Bank00Accesses UMIN"
+             underscore_name="l3_bank00_ic_accesses"
+             units="messages"
+             symbol_name="L3Bank00IcAccesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="L3 Bank 00 IC Hits"
+             description="The total number of hits in L3 Bank 00 from IC cache."
+             data_type="uint64"
+             equation="B 1 READ 2 UMUL $L3Bank00IcAccesses UMIN"
+             underscore_name="l3_bank00_ic_hits"
+             units="messages"
+             symbol_name="L3Bank00IcHits"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3/IC"
+             />
+    <counter name="Sampler Accesses"
+             description="The total number of messages send to samplers."
+             data_type="uint64"
+             equation="A 28 READ"
+             underscore_name="sampler_accesses"
+             units="messages"
+             symbol_name="SamplerAccesses"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler"
+             />
+    <counter name="L3 Bank 01 Accesses"
+             description="The total number of accesses to L3 Bank 01."
+             data_type="uint64"
+             equation="C 1 READ 2 UMUL"
+             underscore_name="l3_bank01_accesses"
+             units="messages"
+             symbol_name="L3Bank01Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU0 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU0."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu0_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu0Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU FPU1 Move Instruction"
+             description="The percentage of time in which execution units were actively processing move instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 20 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_move_fpu1_instruction"
+             units="percent"
+             symbol_name="EuMoveFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="L3 Bank 02 Accesses"
+             description="The total number of accesses to L3 Bank 02."
+             data_type="uint64"
+             equation="B 2 READ 2 UMUL"
+             underscore_name="l3_bank02_accesses"
+             units="messages"
+             symbol_name="L3Bank02Accesses"
+             availability="$SliceMask 0x01 AND"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_hw_unit_type="slice"
+             mdapi_group="L3"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="L3 Total Throughput"
+             description="The total number of GPU memory bytes transferred via L3."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSlicesTotalCount UMUL"
+             equation="$L3Accesses 64 UMUL"
+             underscore_name="l3_total_throughput"
+             units="bytes"
+             symbol_name="L3TotalThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GTI Write Throughput"
+             description="The total number of GPU memory bytes written to GTI."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 32 UMUL"
+             equation="C 7 READ 64 UMUL"
+             underscore_name="gti_write_throughput"
+             units="bytes"
+             symbol_name="GtiWriteThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GTI"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="EU FPU1 Binary Instruction"
+             description="The percentage of time in which execution units were actively processing binary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_binary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuBinaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU FPU1 Ternary Instruction"
+             description="The percentage of time in which execution units were actively processing ternary instructions on FPU1."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_ternary_fpu1_instruction"
+             units="percent"
+             symbol_name="EuTernaryFpu1Instruction"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes/Instructions"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_send_active"
+             units="percent"
+             symbol_name="EuSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x166C0760" />
+        <register type="NOA" address="0x00009888" value="0x1593001E" />
+        <register type="NOA" address="0x00009888" value="0x3F900003" />
+        <register type="NOA" address="0x00009888" value="0x004E8000" />
+        <register type="NOA" address="0x00009888" value="0x0E4E8000" />
+        <register type="NOA" address="0x00009888" value="0x184E8000" />
+        <register type="NOA" address="0x00009888" value="0x1A4E8020" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x006C0051" />
+        <register type="NOA" address="0x00009888" value="0x066C5000" />
+        <register type="NOA" address="0x00009888" value="0x086C5C5D" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5E5F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x186C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x001B4000" />
+        <register type="NOA" address="0x00009888" value="0x061B8000" />
+        <register type="NOA" address="0x00009888" value="0x081BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x101C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1CE000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C0030" />
+        <register type="NOA" address="0x00009888" value="0x004C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C2A00" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0280" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F1500" />
+        <register type="NOA" address="0x00009888" value="0x100F0140" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162C0A00" />
+        <register type="NOA" address="0x00009888" value="0x182C00A0" />
+        <register type="NOA" address="0x00009888" value="0x03933300" />
+        <register type="NOA" address="0x00009888" value="0x05930032" />
+        <register type="NOA" address="0x00009888" value="0x11930000" />
+        <register type="NOA" address="0x00009888" value="0x1B930000" />
+        <register type="NOA" address="0x00009888" value="0x1D900157" />
+        <register type="NOA" address="0x00009888" value="0x1F900158" />
+        <register type="NOA" address="0x00009888" value="0x35900000" />
+        <register type="NOA" address="0x00009888" value="0x19908000" />
+        <register type="NOA" address="0x00009888" value="0x1B908000" />
+        <register type="NOA" address="0x00009888" value="0x1190030F" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900000" />
+        <register type="NOA" address="0x00009888" value="0x55900000" />
+        <register type="NOA" address="0x00009888" value="0x45900021" />
+        <register type="NOA" address="0x00009888" value="0x47900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900000" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x53905555" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x0007FFFA" />
+        <register type="OA" address="0x00002774" value="0x0000FEFE" />
+        <register type="OA" address="0x00002778" value="0x0007FFFA" />
+        <register type="OA" address="0x0000277C" value="0x0000FEFD" />
+        <register type="OA" address="0x00002790" value="0x0007FFFA" />
+        <register type="OA" address="0x00002794" value="0x0000FBEF" />
+        <register type="OA" address="0x00002798" value="0x0007FFFA" />
+        <register type="OA" address="0x0000279C" value="0x0000FBDF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00000003" />
+        <register type="FLEX" address="0x0000E658" value="0x00002001" />
+        <register type="FLEX" address="0x0000E758" value="0x00101100" />
+        <register type="FLEX" address="0x0000E45C" value="0x00201200" />
+        <register type="FLEX" address="0x0000E55C" value="0x00301300" />
+        <register type="FLEX" address="0x0000E65C" value="0x00401400" />
+    </register_config>
+  </set>
+
+  <set name="Metric set HDCAndSF"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="hdc_and_sf"
+       hw_config_guid="730d95dd-7da8-4e1c-ab8d-c0eb1e4c1805"
+       chipset="SKLGT4"
+       symbol_name="HDCAndSF"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Polygon Data Ready"
+             description="The percentage of time in which geometry pipeline output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="poly_data_ready"
+             units="percent"
+             symbol_name="PolyDataReady"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe/Strip-Fans"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="HDC stalled by L3 (s0.ss1)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss1)"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ C 4 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader01_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader01AccessStalledOnL3"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss2)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss2)"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ C 2 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader02_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader02AccessStalledOnL3"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="HDC stalled by L3 (s0.ss0)"
+             description="Percentage of time when HDC has messges to L3, but it's stalled due to lack of credits (s0.ss0)"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ C 6 READ USUB 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_sampler_shader00_access_stalled_on_l3"
+             units="percent"
+             symbol_name="NonSamplerShader00AccessStalledOnL3"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="GPU/Data Port"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x104F0232" />
+        <register type="NOA" address="0x00009888" value="0x124F4640" />
+        <register type="NOA" address="0x00009888" value="0x106C0232" />
+        <register type="NOA" address="0x00009888" value="0x11834400" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x004F1880" />
+        <register type="NOA" address="0x00009888" value="0x024F08BB" />
+        <register type="NOA" address="0x00009888" value="0x044F001B" />
+        <register type="NOA" address="0x00009888" value="0x046C0100" />
+        <register type="NOA" address="0x00009888" value="0x066C000B" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x041B8000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x005B8000" />
+        <register type="NOA" address="0x00009888" value="0x025BC000" />
+        <register type="NOA" address="0x00009888" value="0x045B4000" />
+        <register type="NOA" address="0x00009888" value="0x125C8000" />
+        <register type="NOA" address="0x00009888" value="0x145C8000" />
+        <register type="NOA" address="0x00009888" value="0x165C8000" />
+        <register type="NOA" address="0x00009888" value="0x185C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00A0" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x022CC000" />
+        <register type="NOA" address="0x00009888" value="0x042CC000" />
+        <register type="NOA" address="0x00009888" value="0x062CC000" />
+        <register type="NOA" address="0x00009888" value="0x082CC000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0F828000" />
+        <register type="NOA" address="0x00009888" value="0x0F8305C0" />
+        <register type="NOA" address="0x00009888" value="0x09830000" />
+        <register type="NOA" address="0x00009888" value="0x07830000" />
+        <register type="NOA" address="0x00009888" value="0x1D950080" />
+        <register type="NOA" address="0x00009888" value="0x13928000" />
+        <register type="NOA" address="0x00009888" value="0x0F988000" />
+        <register type="NOA" address="0x00009888" value="0x31904000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x59900001" />
+        <register type="NOA" address="0x00009888" value="0x4B900040" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x10800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000FDFF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_1"
+       hw_config_guid="d9e86d70-462b-462a-851e-fd63e8c13d63"
+       chipset="SKLGT4"
+       symbol_name="L3_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank1 Active"
+             description="The percentage of time in which slice0 L3 bank1 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_active"
+             units="percent"
+             symbol_name="L30Bank1Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 L3 Bank1 Stalled"
+             description="The percentage of time in which slice0 L3 bank1 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank1_stalled"
+             units="percent"
+             symbol_name="L30Bank1Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Active"
+             description="The percentage of time in which slice0 L3 bank0 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_active"
+             units="percent"
+             symbol_name="L30Bank0Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank0 Stalled"
+             description="The percentage of time in which slice0 L3 bank0 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank0_stalled"
+             units="percent"
+             symbol_name="L30Bank0Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C7B40" />
+        <register type="NOA" address="0x00009888" value="0x166C0020" />
+        <register type="NOA" address="0x00009888" value="0x0A603444" />
+        <register type="NOA" address="0x00009888" value="0x0A613400" />
+        <register type="NOA" address="0x00009888" value="0x1A4EA800" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0002" />
+        <register type="NOA" address="0x00009888" value="0x024E8000" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x0C6C5327" />
+        <register type="NOA" address="0x00009888" value="0x0E6C5425" />
+        <register type="NOA" address="0x00009888" value="0x006C2A00" />
+        <register type="NOA" address="0x00009888" value="0x026C285B" />
+        <register type="NOA" address="0x00009888" value="0x046C005C" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1C6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1E6C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0800" />
+        <register type="NOA" address="0x00009888" value="0x0C1BC000" />
+        <register type="NOA" address="0x00009888" value="0x0E1BC000" />
+        <register type="NOA" address="0x00009888" value="0x001B8000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x1C1C003C" />
+        <register type="NOA" address="0x00009888" value="0x121C8000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x10600000" />
+        <register type="NOA" address="0x00009888" value="0x04600000" />
+        <register type="NOA" address="0x00009888" value="0x0C610044" />
+        <register type="NOA" address="0x00009888" value="0x10610000" />
+        <register type="NOA" address="0x00009888" value="0x06610000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02A8" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0154" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x182C00AA" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190FFC0" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900420" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900021" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900400" />
+        <register type="NOA" address="0x00009888" value="0x43900421" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00014002" />
+        <register type="OA" address="0x0000277C" value="0x0000C3FF" />
+        <register type="OA" address="0x00002780" value="0x00010002" />
+        <register type="OA" address="0x00002784" value="0x0000C7FF" />
+        <register type="OA" address="0x00002788" value="0x00004002" />
+        <register type="OA" address="0x0000278C" value="0x0000D3FF" />
+        <register type="OA" address="0x00002790" value="0x00100700" />
+        <register type="OA" address="0x00002794" value="0x0000FF1F" />
+        <register type="OA" address="0x00002798" value="0x00001402" />
+        <register type="OA" address="0x0000279C" value="0x0000FC3F" />
+        <register type="OA" address="0x000027A0" value="0x00001002" />
+        <register type="OA" address="0x000027A4" value="0x0000FC7F" />
+        <register type="OA" address="0x000027A8" value="0x00000402" />
+        <register type="OA" address="0x000027AC" value="0x0000FD3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_2"
+       hw_config_guid="52200424-6ee9-48b3-b7fa-0afcf1975e4d"
+       chipset="SKLGT4"
+       symbol_name="L3_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Stalled"
+             description="The percentage of time in which slice0 L3 bank2 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_stalled"
+             units="percent"
+             symbol_name="L30Bank2Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank2 Active"
+             description="The percentage of time in which slice0 L3 bank2 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank2_active"
+             units="percent"
+             symbol_name="L30Bank2Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C02E0" />
+        <register type="NOA" address="0x00009888" value="0x146C0001" />
+        <register type="NOA" address="0x00009888" value="0x0A623400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x064F4000" />
+        <register type="NOA" address="0x00009888" value="0x026C3324" />
+        <register type="NOA" address="0x00009888" value="0x046C3422" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C0000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C0800" />
+        <register type="NOA" address="0x00009888" value="0x065B4000" />
+        <register type="NOA" address="0x00009888" value="0x1A5C1000" />
+        <register type="NOA" address="0x00009888" value="0x06614000" />
+        <register type="NOA" address="0x00009888" value="0x0C620044" />
+        <register type="NOA" address="0x00009888" value="0x10620000" />
+        <register type="NOA" address="0x00009888" value="0x06620000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C002A" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2CC000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set L3_3"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="l3_3"
+       hw_config_guid="1988315f-0a26-44df-acb0-df7ec86b1456"
+       chipset="SKLGT4"
+       symbol_name="L3_3"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 L3 Bank3 Stalled"
+             description="The percentage of time in which slice0 L3 bank3 is stalled"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_stalled"
+             units="percent"
+             symbol_name="L30Bank3Stalled"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 L3 Bank3 Active"
+             description="The percentage of time in which slice0 L3 bank3 is active"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="l30_bank3_active"
+             units="percent"
+             symbol_name="L30Bank3Active"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GTI/L3"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x126C4E80" />
+        <register type="NOA" address="0x00009888" value="0x146C0000" />
+        <register type="NOA" address="0x00009888" value="0x0A633400" />
+        <register type="NOA" address="0x00009888" value="0x044E8000" />
+        <register type="NOA" address="0x00009888" value="0x064E8000" />
+        <register type="NOA" address="0x00009888" value="0x084E8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4E8000" />
+        <register type="NOA" address="0x00009888" value="0x0C4E8000" />
+        <register type="NOA" address="0x00009888" value="0x026C3321" />
+        <register type="NOA" address="0x00009888" value="0x046C342F" />
+        <register type="NOA" address="0x00009888" value="0x106C0000" />
+        <register type="NOA" address="0x00009888" value="0x1A6C2000" />
+        <register type="NOA" address="0x00009888" value="0x021BC000" />
+        <register type="NOA" address="0x00009888" value="0x041BC000" />
+        <register type="NOA" address="0x00009888" value="0x061B4000" />
+        <register type="NOA" address="0x00009888" value="0x141C8000" />
+        <register type="NOA" address="0x00009888" value="0x161C8000" />
+        <register type="NOA" address="0x00009888" value="0x181C8000" />
+        <register type="NOA" address="0x00009888" value="0x1A1C1800" />
+        <register type="NOA" address="0x00009888" value="0x06604000" />
+        <register type="NOA" address="0x00009888" value="0x0C630044" />
+        <register type="NOA" address="0x00009888" value="0x10630000" />
+        <register type="NOA" address="0x00009888" value="0x06630000" />
+        <register type="NOA" address="0x00009888" value="0x084C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C00AA" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F4000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0055" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190F800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900002" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00100070" />
+        <register type="OA" address="0x00002774" value="0x0000FFF1" />
+        <register type="OA" address="0x00002778" value="0x00028002" />
+        <register type="OA" address="0x0000277C" value="0x000087FF" />
+        <register type="OA" address="0x00002780" value="0x00020002" />
+        <register type="OA" address="0x00002784" value="0x00008FFF" />
+        <register type="OA" address="0x00002788" value="0x00008002" />
+        <register type="OA" address="0x0000278C" value="0x0000A7FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set RasterizerAndPixelBackend"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="rasterizer_and_pixel_backend"
+       hw_config_guid="f1f17ca7-286e-4ae5-9d15-9fccad6c665d"
+       chipset="SKLGT4"
+       symbol_name="RasterizerAndPixelBackend"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Slice0 Pixel Values Ready"
+             description="The percentage of time in which slice0 pixel values are ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_values0_ready"
+             units="percent"
+             symbol_name="PixelValues0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="Slice0 Rasterizer Input Available"
+             description="The percentage of time in which slice0 rasterizer input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_input_available"
+             units="percent"
+             symbol_name="Rasterizer0InputAvailable"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Slice0 Post-EarlyZ Pixel Data Ready"
+             description="The percentage of time in which slice0 post-EarlyZ pixel data is ready (after early Z tests have been applied)"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="pixel_data0_ready"
+             units="percent"
+             symbol_name="PixelData0Ready"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer/Early Depth Test"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Rasterizer Output Ready"
+             description="The percentage of time in which slice0 rasterizer output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="rasterizer0_output_ready"
+             units="percent"
+             symbol_name="Rasterizer0OutputReady"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Rasterizer"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Slice0 PS Output Available"
+             description="The percentage of time in which slice0 PS output is available"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_output0_available"
+             units="percent"
+             symbol_name="PSOutput0Available"
+             availability="$SliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/3D Pipe"
+             mdapi_hw_unit_type="slice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x102F3800" />
+        <register type="NOA" address="0x00009888" value="0x144D0500" />
+        <register type="NOA" address="0x00009888" value="0x120D03C0" />
+        <register type="NOA" address="0x00009888" value="0x140D03CF" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0004" />
+        <register type="NOA" address="0x00009888" value="0x0C4E4000" />
+        <register type="NOA" address="0x00009888" value="0x042F0480" />
+        <register type="NOA" address="0x00009888" value="0x082F0000" />
+        <register type="NOA" address="0x00009888" value="0x022F0000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0090" />
+        <register type="NOA" address="0x00009888" value="0x064D0027" />
+        <register type="NOA" address="0x00009888" value="0x004D0000" />
+        <register type="NOA" address="0x00009888" value="0x000D0D40" />
+        <register type="NOA" address="0x00009888" value="0x020D803F" />
+        <register type="NOA" address="0x00009888" value="0x040D8023" />
+        <register type="NOA" address="0x00009888" value="0x100D0000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x020F0010" />
+        <register type="NOA" address="0x00009888" value="0x000F0000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0050" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41901400" />
+        <register type="NOA" address="0x00009888" value="0x43901485" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900001" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x30800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x0000EFFF" />
+        <register type="OA" address="0x00002778" value="0x00006000" />
+        <register type="OA" address="0x0000277C" value="0x0000F3FF" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set Sampler"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="sampler"
+       hw_config_guid="00a9e0fb-3d2e-4405-852c-dce6334ffb3b"
+       chipset="SKLGT4"
+       symbol_name="Sampler"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice2 Input Available"
+             description="The percentage of time in which slice0 subslice2 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_input_available"
+             units="percent"
+             symbol_name="Sampler02InputAvailable"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Slice0 Subslice0 Input Available"
+             description="The percentage of time in which slice0 subslice0 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_input_available"
+             units="percent"
+             symbol_name="Sampler00InputAvailable"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="Slice0 Subslice2 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice2 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler02_output_ready"
+             units="percent"
+             symbol_name="Sampler02OutputReady"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Slice0 Subslice1 Input Available"
+             description="The percentage of time in which slice0 subslice1 sampler input is available"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_input_available"
+             units="percent"
+             symbol_name="Sampler01InputAvailable"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Slice0 Subslice0 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice0 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler00_output_ready"
+             units="percent"
+             symbol_name="Sampler00OutputReady"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Slice0 Subslice1 Sampler Output Ready"
+             description="The percentage of time in which slice0 subslice1 sampler output is ready"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="sampler01_output_ready"
+             units="percent"
+             symbol_name="Sampler01OutputReady"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Sampler"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x14152C00" />
+        <register type="NOA" address="0x00009888" value="0x16150005" />
+        <register type="NOA" address="0x00009888" value="0x121600A0" />
+        <register type="NOA" address="0x00009888" value="0x14352C00" />
+        <register type="NOA" address="0x00009888" value="0x16350005" />
+        <register type="NOA" address="0x00009888" value="0x123600A0" />
+        <register type="NOA" address="0x00009888" value="0x14552C00" />
+        <register type="NOA" address="0x00009888" value="0x16550005" />
+        <register type="NOA" address="0x00009888" value="0x125600A0" />
+        <register type="NOA" address="0x00009888" value="0x062F6000" />
+        <register type="NOA" address="0x00009888" value="0x022F2000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C0050" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0010" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0350" />
+        <register type="NOA" address="0x00009888" value="0x0C0FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F00DA" />
+        <register type="NOA" address="0x00009888" value="0x182C0028" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x022DC000" />
+        <register type="NOA" address="0x00009888" value="0x042D4000" />
+        <register type="NOA" address="0x00009888" value="0x0C138000" />
+        <register type="NOA" address="0x00009888" value="0x0E132000" />
+        <register type="NOA" address="0x00009888" value="0x0413C000" />
+        <register type="NOA" address="0x00009888" value="0x1C140018" />
+        <register type="NOA" address="0x00009888" value="0x0C157000" />
+        <register type="NOA" address="0x00009888" value="0x0E150078" />
+        <register type="NOA" address="0x00009888" value="0x10150000" />
+        <register type="NOA" address="0x00009888" value="0x04162180" />
+        <register type="NOA" address="0x00009888" value="0x02160000" />
+        <register type="NOA" address="0x00009888" value="0x04174000" />
+        <register type="NOA" address="0x00009888" value="0x0233A000" />
+        <register type="NOA" address="0x00009888" value="0x04333000" />
+        <register type="NOA" address="0x00009888" value="0x14348000" />
+        <register type="NOA" address="0x00009888" value="0x16348000" />
+        <register type="NOA" address="0x00009888" value="0x02357870" />
+        <register type="NOA" address="0x00009888" value="0x10350000" />
+        <register type="NOA" address="0x00009888" value="0x04360043" />
+        <register type="NOA" address="0x00009888" value="0x02360000" />
+        <register type="NOA" address="0x00009888" value="0x04371000" />
+        <register type="NOA" address="0x00009888" value="0x0E538000" />
+        <register type="NOA" address="0x00009888" value="0x00538000" />
+        <register type="NOA" address="0x00009888" value="0x06533000" />
+        <register type="NOA" address="0x00009888" value="0x1C540020" />
+        <register type="NOA" address="0x00009888" value="0x12548000" />
+        <register type="NOA" address="0x00009888" value="0x0E557000" />
+        <register type="NOA" address="0x00009888" value="0x00557800" />
+        <register type="NOA" address="0x00009888" value="0x10550000" />
+        <register type="NOA" address="0x00009888" value="0x06560043" />
+        <register type="NOA" address="0x00009888" value="0x02560000" />
+        <register type="NOA" address="0x00009888" value="0x06571000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900060" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900842" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900060" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x70800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+        <register type="OA" address="0x00002770" value="0x0000C000" />
+        <register type="OA" address="0x00002774" value="0x0000E7FF" />
+        <register type="OA" address="0x00002778" value="0x00003000" />
+        <register type="OA" address="0x0000277C" value="0x0000F9FF" />
+        <register type="OA" address="0x00002780" value="0x00000C00" />
+        <register type="OA" address="0x00002784" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_1"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_1"
+       hw_config_guid="13dcc50a-7ec0-409b-99d6-a3f932cedcb3"
+       chipset="SKLGT4"
+       symbol_name="TDL_1"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice0 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice0 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 1 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread00_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread00ReadyForDispatch"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="PS Thread Ready For Dispatch on Slice0 Subslice2 Thread Dispatcher"
+             description="The percentage of time in which PS thread is ready for dispatch on slice0 subslice2 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_thread02_ready_for_dispatch"
+             units="percent"
+             symbol_name="PSThread02ReadyForDispatch"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Non-PS Thread Ready For Dispatch on Slice0 Subslice1 Thread Dispatcher"
+             description="The percentage of time in which non-PS thread is ready for dispatch on slice0 subslice1 thread dispatcher"
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="non_ps_thread01_ready_for_dispatch"
+             units="percent"
+             symbol_name="NonPSThread01ReadyForDispatch"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12120000" />
+        <register type="NOA" address="0x00009888" value="0x12320000" />
+        <register type="NOA" address="0x00009888" value="0x12520000" />
+        <register type="NOA" address="0x00009888" value="0x002F8000" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0015" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F03A0" />
+        <register type="NOA" address="0x00009888" value="0x0C0FF000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0095" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2D4000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x02108000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x02118000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x02121880" />
+        <register type="NOA" address="0x00009888" value="0x041219B5" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x02134000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x0C308000" />
+        <register type="NOA" address="0x00009888" value="0x0E304000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x0C318000" />
+        <register type="NOA" address="0x00009888" value="0x0E314000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x0C321A80" />
+        <register type="NOA" address="0x00009888" value="0x0E320033" />
+        <register type="NOA" address="0x00009888" value="0x06320031" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x0C334000" />
+        <register type="NOA" address="0x00009888" value="0x0E331000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0E508000" />
+        <register type="NOA" address="0x00009888" value="0x00508000" />
+        <register type="NOA" address="0x00009888" value="0x02504000" />
+        <register type="NOA" address="0x00009888" value="0x0E518000" />
+        <register type="NOA" address="0x00009888" value="0x00518000" />
+        <register type="NOA" address="0x00009888" value="0x02514000" />
+        <register type="NOA" address="0x00009888" value="0x0E521880" />
+        <register type="NOA" address="0x00009888" value="0x00521A80" />
+        <register type="NOA" address="0x00009888" value="0x02520033" />
+        <register type="NOA" address="0x00009888" value="0x0E534000" />
+        <register type="NOA" address="0x00009888" value="0x00534000" />
+        <register type="NOA" address="0x00009888" value="0x02531000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900800" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900062" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900C00" />
+        <register type="NOA" address="0x00009888" value="0x43900003" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00000002" />
+        <register type="OA" address="0x00002774" value="0x00007FFF" />
+        <register type="OA" address="0x00002778" value="0x00000000" />
+        <register type="OA" address="0x0000277C" value="0x00009FFF" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000EFFF" />
+        <register type="OA" address="0x00002788" value="0x00000000" />
+        <register type="OA" address="0x0000278C" value="0x0000F3FF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FDFF" />
+        <register type="OA" address="0x00002798" value="0x00000000" />
+        <register type="OA" address="0x0000279C" value="0x0000FE7F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Metric set TDL_2"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="tdl_2"
+       hw_config_guid="97875e21-6624-4aee-9191-682feb3eae21"
+       chipset="SKLGT4"
+       symbol_name="TDL_2"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Send Pipe Active"
+             description="The percentage of time in which EU send pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 12 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_send_active"
+             units="percent"
+             symbol_name="VsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 11 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu1_active"
+             units="percent"
+             symbol_name="VsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GS Threads Dispatched"
+             description="The total number of geometry shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 5 READ"
+             underscore_name="gs_threads"
+             units="threads"
+             symbol_name="GsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Geometry Shader"
+             />
+    <counter name="Early Hi-Depth Test Fails"
+             description="The total number of pixels dropped on early hierarchical depth test."
+             data_type="uint64"
+             equation="A 22 READ 4 UMUL"
+             underscore_name="hi_depth_test_fails"
+             units="pixels"
+             symbol_name="HiDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Hi-Depth Test"
+             />
+    <counter name="FS Both FPU Active"
+             description="The percentage of time in which fragment shaders were processed actively on the both FPUs."
+             data_type="float"
+             max_equation="100"
+             equation="A 18 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_eu_both_fpu_active"
+             units="percent"
+             symbol_name="PsEuBothFpuActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Frame Batch Draw"
+             mdapi_group="3D Pipe/Fragment Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS Threads Dispatched"
+             description="The total number of vertex shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 1 READ"
+             underscore_name="vs_threads"
+             units="threads"
+             symbol_name="VsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Vertex Shader"
+             />
+    <counter name="FS Threads Dispatched"
+             description="The total number of fragment shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 6 READ"
+             underscore_name="ps_threads"
+             units="threads"
+             symbol_name="PsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Fragment Shader"
+             />
+    <counter name="Shader Barrier Messages"
+             description="The total number of shader barrier messages."
+             data_type="uint64"
+             equation="A 35 READ"
+             underscore_name="shader_barriers"
+             units="messages"
+             symbol_name="ShaderBarriers"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Barrier"
+             />
+    <counter name="Sampler Texels"
+             description="The total number of texels seen on input (with 2x2 accuracy) in all sampler units."
+             data_type="uint64"
+             equation="A 28 READ 4 UMUL"
+             underscore_name="sampler_texels"
+             units="texels"
+             symbol_name="SamplerTexels"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Input"
+             />
+    <counter name="Pixels Failing Tests"
+             description="The total number of pixels dropped on post-FS alpha, stencil, or depth tests."
+             data_type="uint64"
+             equation="A 25 READ 4 UMUL"
+             underscore_name="pixels_failing_post_ps_tests"
+             units="pixels"
+             symbol_name="PixelsFailingPostPsTests"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 7 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort0"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice1 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice1 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 2 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header01_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader01ReadyPort1"
+             availability="$SubsliceMask 0x2 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Sampler Texels Misses"
+             description="The total number of texels lookups (with 2x2 accuracy) that missed L1 sampler cache."
+             data_type="uint64"
+             equation="A 29 READ 4 UMUL"
+             underscore_name="sampler_texel_misses"
+             units="texels"
+             symbol_name="SamplerTexelMisses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Batch Frame Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="Sampler/Sampler Cache"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="SLM Bytes Read"
+             description="The total number of GPU memory bytes read from shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ 64 UMUL"
+             underscore_name="slm_bytes_read"
+             units="bytes"
+             symbol_name="SlmBytesRead"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 16 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu1_active"
+             units="percent"
+             symbol_name="PsFpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="PS Send Pipeline Active"
+             description="The percentage of time in which EU send pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 17 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_send_active"
+             units="percent"
+             symbol_name="PsSendActive"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a vertex shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 10 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vs_fpu0_active"
+             units="percent"
+             symbol_name="VsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Vertex Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 4 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort1"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice2 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice2 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 3 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header02_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader02ReadyPort0"
+             availability="$SubsliceMask 0x4 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Rasterized Pixels"
+             description="The total number of rasterized pixels."
+             data_type="uint64"
+             equation="A 21 READ 4 UMUL"
+             underscore_name="rasterized_pixels"
+             units="pixels"
+             symbol_name="RasterizedPixels"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer"
+             />
+    <counter name="PS FPU0 Pipe Active"
+             description="The percentage of time in which EU FPU0 pipeline was actively processing a pixel shader instruction."
+             data_type="float"
+             max_equation="100"
+             equation="A 15 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="ps_fpu0_active"
+             units="percent"
+             symbol_name="PsFpu0Active"
+             semantic_type="duration"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pixel Shader"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="DS Threads Dispatched"
+             description="The total number of domain shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 3 READ"
+             underscore_name="ds_threads"
+             units="threads"
+             symbol_name="DsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Domain Shader"
+             />
+    <counter name="Samples Written"
+             description="The total number of samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 26 READ 4 UMUL"
+             underscore_name="samples_written"
+             units="pixels"
+             symbol_name="SamplesWritten"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Blended"
+             description="The total number of blended samples or pixels written to all render targets."
+             data_type="uint64"
+             equation="A 27 READ 4 UMUL"
+             underscore_name="samples_blended"
+             units="pixels"
+             symbol_name="SamplesBlended"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Output Merger"
+             />
+    <counter name="Early Depth Test Fails"
+             description="The total number of pixels dropped on early depth test."
+             data_type="uint64"
+             equation="A 23 READ 4 UMUL"
+             underscore_name="early_depth_test_fails"
+             units="pixels"
+             symbol_name="EarlyDepthTestFails"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Rasterizer/Early Depth Test"
+             />
+    <counter name="Shader Memory Accesses"
+             description="The total number of shader memory accesses to L3."
+             data_type="uint64"
+             equation="A 32 READ"
+             underscore_name="shader_memory_accesses"
+             units="messages"
+             symbol_name="ShaderMemoryAccesses"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port"
+             />
+    <counter name="HS Threads Dispatched"
+             description="The total number of hull shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 2 READ"
+             underscore_name="hs_threads"
+             units="threads"
+             symbol_name="HsThreads"
+             semantic_type="event"
+             mdapi_supported_apis="OGL4 IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Hull Shader"
+             />
+    <counter name="SLM Bytes Written"
+             description="The total number of GPU memory bytes written into shared local memory."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 31 READ 64 UMUL"
+             underscore_name="slm_bytes_written"
+             units="bytes"
+             symbol_name="SlmBytesWritten"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_group="L3/Data Port/SLM"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="L3 Shader Throughput"
+             description="The total number of GPU memory bytes transferred between shaders and L3 caches w/o URB."
+             data_type="uint64"
+             max_equation="$GpuCoreClocks 64 UMUL $EuSubslicesTotalCount UMUL"
+             equation="A 30 READ  A 31 READ $ShaderMemoryAccesses 64 UMUL UADD UADD"
+             underscore_name="l3_shader_throughput"
+             units="bytes"
+             symbol_name="L3ShaderThroughput"
+             semantic_type="throughput"
+             mdapi_supported_apis="OGL4 OCL IO BB"
+             mdapi_usage_flags="Tier2 Frame Batch Draw"
+             mdapi_group="L3/Data Port"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="Samples Killed in FS"
+             description="The total number of samples or pixels dropped in fragment shaders."
+             data_type="uint64"
+             equation="A 24 READ 4 UMUL"
+             underscore_name="samples_killed_in_ps"
+             units="pixels"
+             symbol_name="SamplesKilledInPs"
+             semantic_type="event"
+             mdapi_supported_apis="OGL IO BB"
+             mdapi_usage_flags="Tier4 Overview Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="3D Pipe/Fragment Shader"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 0"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 0"
+             data_type="float"
+             max_equation="100"
+             equation="C 5 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port0"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort0"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <counter name="Shader Atomic Memory Accesses"
+             description="The total number of shader atomic memory accesses."
+             data_type="uint64"
+             equation="A 34 READ"
+             underscore_name="shader_atomics"
+             units="messages"
+             symbol_name="ShaderAtomics"
+             semantic_type="event"
+             mdapi_supported_apis="OGL OCL IO BB"
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="L3/Data Port/Atomics"
+             />
+    <counter name="Thread Header Ready on Slice0 Subslice0 Thread Dispatcher Port 1"
+             description="The percentage of time in which thread header is ready on slice0 subslice0 thread dispatcher port 1"
+             data_type="float"
+             max_equation="100"
+             equation="C 6 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="thread_header00_ready_port1"
+             units="percent"
+             symbol_name="ThreadHeader00ReadyPort1"
+             availability="$SubsliceMask 0x1 AND"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview Frame Batch Draw"
+             mdapi_group="GPU/Thread Dispatcher"
+             mdapi_hw_unit_type="subslice"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x12124D60" />
+        <register type="NOA" address="0x00009888" value="0x12322E60" />
+        <register type="NOA" address="0x00009888" value="0x12524D60" />
+        <register type="NOA" address="0x00009888" value="0x022F3000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0014" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x0C0FE000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F0097" />
+        <register type="NOA" address="0x00009888" value="0x082C8000" />
+        <register type="NOA" address="0x00009888" value="0x0A2C8000" />
+        <register type="NOA" address="0x00009888" value="0x002D8000" />
+        <register type="NOA" address="0x00009888" value="0x062D4000" />
+        <register type="NOA" address="0x00009888" value="0x0410C000" />
+        <register type="NOA" address="0x00009888" value="0x0411C000" />
+        <register type="NOA" address="0x00009888" value="0x04121FB7" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x04135000" />
+        <register type="NOA" address="0x00009888" value="0x00308000" />
+        <register type="NOA" address="0x00009888" value="0x06304000" />
+        <register type="NOA" address="0x00009888" value="0x00318000" />
+        <register type="NOA" address="0x00009888" value="0x06314000" />
+        <register type="NOA" address="0x00009888" value="0x00321B80" />
+        <register type="NOA" address="0x00009888" value="0x0632003F" />
+        <register type="NOA" address="0x00009888" value="0x00334000" />
+        <register type="NOA" address="0x00009888" value="0x06331000" />
+        <register type="NOA" address="0x00009888" value="0x0250C000" />
+        <register type="NOA" address="0x00009888" value="0x0251C000" />
+        <register type="NOA" address="0x00009888" value="0x02521FB7" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x02535000" />
+        <register type="NOA" address="0x00009888" value="0x1190FC00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x51900000" />
+        <register type="NOA" address="0x00009888" value="0x41900800" />
+        <register type="NOA" address="0x00009888" value="0x43900063" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900040" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0x00800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x00800000" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00010003" />
+        <register type="FLEX" address="0x0000E658" value="0x00012011" />
+        <register type="FLEX" address="0x0000E758" value="0x00015014" />
+        <register type="FLEX" address="0x0000E45C" value="0x00051050" />
+        <register type="FLEX" address="0x0000E55C" value="0x00053052" />
+        <register type="FLEX" address="0x0000E65C" value="0x00055054" />
+    </register_config>
+  </set>
+
+  <set name="Compute Metrics Extra Gen9"
+       mdapi_supported_apis="OGL4 OCL IO BB"
+       underscore_name="compute_extra"
+       hw_config_guid="a5aa857d-e8f0-4dfa-8981-ce340fa748fd"
+       chipset="SKLGT4"
+       symbol_name="ComputeExtra"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU FPU1 Pipe Active"
+             description="The percentage of time in which EU FPU1 pipeline was actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="fpu1_active"
+             units="percent"
+             symbol_name="Fpu1Active"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x121203E0" />
+        <register type="NOA" address="0x00009888" value="0x123203E0" />
+        <register type="NOA" address="0x00009888" value="0x125203E0" />
+        <register type="NOA" address="0x00009888" value="0x129203E0" />
+        <register type="NOA" address="0x00009888" value="0x12B203E0" />
+        <register type="NOA" address="0x00009888" value="0x12D203E0" />
+        <register type="NOA" address="0x00009888" value="0x131203E0" />
+        <register type="NOA" address="0x00009888" value="0x133203E0" />
+        <register type="NOA" address="0x00009888" value="0x135203E0" />
+        <register type="NOA" address="0x00009888" value="0x1A4EF000" />
+        <register type="NOA" address="0x00009888" value="0x1C4E0003" />
+        <register type="NOA" address="0x00009888" value="0x024EC000" />
+        <register type="NOA" address="0x00009888" value="0x044EC000" />
+        <register type="NOA" address="0x00009888" value="0x064EC000" />
+        <register type="NOA" address="0x00009888" value="0x022F4000" />
+        <register type="NOA" address="0x00009888" value="0x0C4C02A0" />
+        <register type="NOA" address="0x00009888" value="0x084CA000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C0042" />
+        <register type="NOA" address="0x00009888" value="0x0C0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E0DA000" />
+        <register type="NOA" address="0x00009888" value="0x000D8000" />
+        <register type="NOA" address="0x00009888" value="0x020DA000" />
+        <register type="NOA" address="0x00009888" value="0x040DA000" />
+        <register type="NOA" address="0x00009888" value="0x060D2000" />
+        <register type="NOA" address="0x00009888" value="0x100F0150" />
+        <register type="NOA" address="0x00009888" value="0x0C0F5000" />
+        <register type="NOA" address="0x00009888" value="0x0E0F006D" />
+        <register type="NOA" address="0x00009888" value="0x182C00A8" />
+        <register type="NOA" address="0x00009888" value="0x022C8000" />
+        <register type="NOA" address="0x00009888" value="0x042C8000" />
+        <register type="NOA" address="0x00009888" value="0x062C8000" />
+        <register type="NOA" address="0x00009888" value="0x0C2C8000" />
+        <register type="NOA" address="0x00009888" value="0x042D8000" />
+        <register type="NOA" address="0x00009888" value="0x06104000" />
+        <register type="NOA" address="0x00009888" value="0x06114000" />
+        <register type="NOA" address="0x00009888" value="0x06120033" />
+        <register type="NOA" address="0x00009888" value="0x00120000" />
+        <register type="NOA" address="0x00009888" value="0x06131000" />
+        <register type="NOA" address="0x00009888" value="0x04308000" />
+        <register type="NOA" address="0x00009888" value="0x04318000" />
+        <register type="NOA" address="0x00009888" value="0x04321980" />
+        <register type="NOA" address="0x00009888" value="0x00320000" />
+        <register type="NOA" address="0x00009888" value="0x04334000" />
+        <register type="NOA" address="0x00009888" value="0x04504000" />
+        <register type="NOA" address="0x00009888" value="0x04514000" />
+        <register type="NOA" address="0x00009888" value="0x04520033" />
+        <register type="NOA" address="0x00009888" value="0x00520000" />
+        <register type="NOA" address="0x00009888" value="0x04531000" />
+        <register type="NOA" address="0x00009888" value="0x1ACEF000" />
+        <register type="NOA" address="0x00009888" value="0x1CCE0003" />
+        <register type="NOA" address="0x00009888" value="0x00AF8000" />
+        <register type="NOA" address="0x00009888" value="0x0CCC02A0" />
+        <register type="NOA" address="0x00009888" value="0x0ACC0001" />
+        <register type="NOA" address="0x00009888" value="0x0C8D8000" />
+        <register type="NOA" address="0x00009888" value="0x0E8DA000" />
+        <register type="NOA" address="0x00009888" value="0x008D8000" />
+        <register type="NOA" address="0x00009888" value="0x028DA000" />
+        <register type="NOA" address="0x00009888" value="0x108F0150" />
+        <register type="NOA" address="0x00009888" value="0x0C8FB000" />
+        <register type="NOA" address="0x00009888" value="0x0E8F0001" />
+        <register type="NOA" address="0x00009888" value="0x18AC00A8" />
+        <register type="NOA" address="0x00009888" value="0x06AC8000" />
+        <register type="NOA" address="0x00009888" value="0x02AD4000" />
+        <register type="NOA" address="0x00009888" value="0x02908000" />
+        <register type="NOA" address="0x00009888" value="0x02918000" />
+        <register type="NOA" address="0x00009888" value="0x02921980" />
+        <register type="NOA" address="0x00009888" value="0x00920000" />
+        <register type="NOA" address="0x00009888" value="0x02934000" />
+        <register type="NOA" address="0x00009888" value="0x02B04000" />
+        <register type="NOA" address="0x00009888" value="0x02B14000" />
+        <register type="NOA" address="0x00009888" value="0x02B20033" />
+        <register type="NOA" address="0x00009888" value="0x00B20000" />
+        <register type="NOA" address="0x00009888" value="0x02B31000" />
+        <register type="NOA" address="0x00009888" value="0x00D08000" />
+        <register type="NOA" address="0x00009888" value="0x00D18000" />
+        <register type="NOA" address="0x00009888" value="0x00D21980" />
+        <register type="NOA" address="0x00009888" value="0x00D34000" />
+        <register type="NOA" address="0x00009888" value="0x072F8000" />
+        <register type="NOA" address="0x00009888" value="0x0D4C0100" />
+        <register type="NOA" address="0x00009888" value="0x0D0D8000" />
+        <register type="NOA" address="0x00009888" value="0x0F0DA000" />
+        <register type="NOA" address="0x00009888" value="0x110F01B0" />
+        <register type="NOA" address="0x00009888" value="0x192C0080" />
+        <register type="NOA" address="0x00009888" value="0x0F2D4000" />
+        <register type="NOA" address="0x00009888" value="0x0F108000" />
+        <register type="NOA" address="0x00009888" value="0x0F118000" />
+        <register type="NOA" address="0x00009888" value="0x0F121980" />
+        <register type="NOA" address="0x00009888" value="0x01120000" />
+        <register type="NOA" address="0x00009888" value="0x0F134000" />
+        <register type="NOA" address="0x00009888" value="0x0F304000" />
+        <register type="NOA" address="0x00009888" value="0x0F314000" />
+        <register type="NOA" address="0x00009888" value="0x0F320033" />
+        <register type="NOA" address="0x00009888" value="0x01320000" />
+        <register type="NOA" address="0x00009888" value="0x0F331000" />
+        <register type="NOA" address="0x00009888" value="0x0D508000" />
+        <register type="NOA" address="0x00009888" value="0x0D518000" />
+        <register type="NOA" address="0x00009888" value="0x0D521980" />
+        <register type="NOA" address="0x00009888" value="0x01520000" />
+        <register type="NOA" address="0x00009888" value="0x0D534000" />
+        <register type="NOA" address="0x00009888" value="0x1190FF80" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900C00" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+        <register type="NOA" address="0x00009888" value="0x4B900002" />
+        <register type="NOA" address="0x00009888" value="0x59900000" />
+        <register type="NOA" address="0x00009888" value="0x51901100" />
+        <register type="NOA" address="0x00009888" value="0x41901000" />
+        <register type="NOA" address="0x00009888" value="0x43901423" />
+        <register type="NOA" address="0x00009888" value="0x53903331" />
+        <register type="NOA" address="0x00009888" value="0x45900044" />
+    </register_config>
+  </set>
+
+  <set name="Media Vme Pipe Gen9"
+       mdapi_supported_apis="MEDIA IO BB"
+       underscore_name="vme_pipe"
+       hw_config_guid="0e8d8b86-4ee7-4cdd-aaaa-58adc92cb29e"
+       chipset="SKLGT4"
+       symbol_name="VMEPipe"
+       >
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="EU Active"
+             description="The percentage of time in which the Execution Units were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 7 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_active"
+             units="percent"
+             symbol_name="EuActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Both FPU Pipes Active"
+             description="The percentage of time in which both EU FPU pipelines were actively processing."
+             data_type="float"
+             max_equation="100"
+             equation="A 9 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_fpu_both_active"
+             units="percent"
+             symbol_name="EuFpuBothActive"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Overview System Frame Batch Draw"
+             mdapi_group="EU Array/Pipes"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL  GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="CS Threads Dispatched"
+             description="The total number of compute shader hardware threads dispatched."
+             data_type="uint64"
+             equation="A 4 READ"
+             underscore_name="cs_threads"
+             units="threads"
+             symbol_name="CsThreads"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier3 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="EU Array/Compute Shader"
+             />
+    <counter name="EU Thread Occupancy"
+             description="The percentage of time in which hardware threads occupied EUs."
+             data_type="float"
+             max_equation="100"
+             equation="8 A 10 READ FMUL $EuThreadsCount FDIV $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_thread_occupancy"
+             units="percent"
+             symbol_name="EuThreadOccupancy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="EU Stall"
+             description="The percentage of time in which the Execution Units were stalled."
+             data_type="float"
+             max_equation="100"
+             equation="A 8 READ $EuCoresTotalCount UDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="eu_stall"
+             units="percent"
+             symbol_name="EuStall"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier2 Overview System Frame Batch Draw"
+             mdapi_group="EU Array"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="VME Busy"
+             description="The percentage of time in which VME (IME or CRE) was actively processing data."
+             data_type="float"
+             max_equation="100"
+             equation="B 0 READ B 3 READ FADD 2 FDIV 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="vme_busy"
+             units="percent"
+             symbol_name="VMEBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Overview System Batch Tier2"
+             mdapi_group="VME Pipe"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="GPU Busy"
+             description="The percentage of time in which the GPU has been processing GPU commands."
+             data_type="float"
+             max_equation="100"
+             equation="A 0 READ 100 UMUL $GpuCoreClocks FDIV"
+             underscore_name="gpu_busy"
+             units="percent"
+             symbol_name="GpuBusy"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x141A5800" />
+        <register type="NOA" address="0x00009888" value="0x161A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12180240" />
+        <register type="NOA" address="0x00009888" value="0x14180002" />
+        <register type="NOA" address="0x00009888" value="0x149A5800" />
+        <register type="NOA" address="0x00009888" value="0x169A00C0" />
+        <register type="NOA" address="0x00009888" value="0x12980240" />
+        <register type="NOA" address="0x00009888" value="0x14980002" />
+        <register type="NOA" address="0x00009888" value="0x1A4E3FC0" />
+        <register type="NOA" address="0x00009888" value="0x002F1000" />
+        <register type="NOA" address="0x00009888" value="0x022F8000" />
+        <register type="NOA" address="0x00009888" value="0x042F3000" />
+        <register type="NOA" address="0x00009888" value="0x004C4000" />
+        <register type="NOA" address="0x00009888" value="0x0A4C9500" />
+        <register type="NOA" address="0x00009888" value="0x0C4C002A" />
+        <register type="NOA" address="0x00009888" value="0x000D2000" />
+        <register type="NOA" address="0x00009888" value="0x060D8000" />
+        <register type="NOA" address="0x00009888" value="0x080DA000" />
+        <register type="NOA" address="0x00009888" value="0x0A0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C0F0400" />
+        <register type="NOA" address="0x00009888" value="0x0E0F5500" />
+        <register type="NOA" address="0x00009888" value="0x100F0015" />
+        <register type="NOA" address="0x00009888" value="0x002C8000" />
+        <register type="NOA" address="0x00009888" value="0x0E2C8000" />
+        <register type="NOA" address="0x00009888" value="0x162CAA00" />
+        <register type="NOA" address="0x00009888" value="0x182C000A" />
+        <register type="NOA" address="0x00009888" value="0x04193000" />
+        <register type="NOA" address="0x00009888" value="0x081A28C1" />
+        <register type="NOA" address="0x00009888" value="0x001A0000" />
+        <register type="NOA" address="0x00009888" value="0x00133000" />
+        <register type="NOA" address="0x00009888" value="0x0613C000" />
+        <register type="NOA" address="0x00009888" value="0x0813F000" />
+        <register type="NOA" address="0x00009888" value="0x00172000" />
+        <register type="NOA" address="0x00009888" value="0x06178000" />
+        <register type="NOA" address="0x00009888" value="0x0817A000" />
+        <register type="NOA" address="0x00009888" value="0x00180037" />
+        <register type="NOA" address="0x00009888" value="0x06180940" />
+        <register type="NOA" address="0x00009888" value="0x08180000" />
+        <register type="NOA" address="0x00009888" value="0x02180000" />
+        <register type="NOA" address="0x00009888" value="0x04183000" />
+        <register type="NOA" address="0x00009888" value="0x04AFC000" />
+        <register type="NOA" address="0x00009888" value="0x06AF3000" />
+        <register type="NOA" address="0x00009888" value="0x0ACC4000" />
+        <register type="NOA" address="0x00009888" value="0x0CCC0015" />
+        <register type="NOA" address="0x00009888" value="0x0A8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0C8DA000" />
+        <register type="NOA" address="0x00009888" value="0x0E8F4000" />
+        <register type="NOA" address="0x00009888" value="0x108F0015" />
+        <register type="NOA" address="0x00009888" value="0x16ACA000" />
+        <register type="NOA" address="0x00009888" value="0x18AC000A" />
+        <register type="NOA" address="0x00009888" value="0x06993000" />
+        <register type="NOA" address="0x00009888" value="0x0C9A28C1" />
+        <register type="NOA" address="0x00009888" value="0x009A0000" />
+        <register type="NOA" address="0x00009888" value="0x0A93F000" />
+        <register type="NOA" address="0x00009888" value="0x0C93F000" />
+        <register type="NOA" address="0x00009888" value="0x0A97A000" />
+        <register type="NOA" address="0x00009888" value="0x0C97A000" />
+        <register type="NOA" address="0x00009888" value="0x0A980977" />
+        <register type="NOA" address="0x00009888" value="0x08980000" />
+        <register type="NOA" address="0x00009888" value="0x04980000" />
+        <register type="NOA" address="0x00009888" value="0x06983000" />
+        <register type="NOA" address="0x00009888" value="0x119000FF" />
+        <register type="NOA" address="0x00009888" value="0x51900010" />
+        <register type="NOA" address="0x00009888" value="0x41900060" />
+        <register type="NOA" address="0x00009888" value="0x55900111" />
+        <register type="NOA" address="0x00009888" value="0x45900C00" />
+        <register type="NOA" address="0x00009888" value="0x47900821" />
+        <register type="NOA" address="0x00009888" value="0x57900000" />
+        <register type="NOA" address="0x00009888" value="0x49900002" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0x30800000" />
+        <register type="OA" address="0x00002770" value="0x00100030" />
+        <register type="OA" address="0x00002774" value="0x0000FFF9" />
+        <register type="OA" address="0x00002778" value="0x00000002" />
+        <register type="OA" address="0x0000277C" value="0x0000FFFC" />
+        <register type="OA" address="0x00002780" value="0x00000002" />
+        <register type="OA" address="0x00002784" value="0x0000FFF3" />
+        <register type="OA" address="0x00002788" value="0x00100180" />
+        <register type="OA" address="0x0000278C" value="0x0000FFCF" />
+        <register type="OA" address="0x00002790" value="0x00000002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00000002" />
+        <register type="OA" address="0x0000279C" value="0x0000FF3F" />
+    </register_config>
+    <register_config type="FLEX">
+        <register type="FLEX" address="0x0000E458" value="0x00005004" />
+        <register type="FLEX" address="0x0000E558" value="0x00008003" />
+    </register_config>
+  </set>
+
+  <set name="MDAPI testing set Gen9"
+       mdapi_supported_apis="OGL OCL IO BB"
+       underscore_name="test_oa"
+       hw_config_guid="882fa433-1f4a-4a67-a962-c741888fe5f5"
+       chipset="SKLGT4"
+       symbol_name="TestOa"
+       >
+    <counter name="TestCounter7"
+             description="HW test counter 7. Factor: 0.666"
+             data_type="uint64"
+             equation="B 7 READ"
+             underscore_name="counter7"
+             units="events"
+             symbol_name="Counter7"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Time Elapsed"
+             description="Time elapsed on the GPU during the measurement."
+             data_type="uint64"
+             equation="GPU_TIME 0 READ 1000000000 UMUL $GpuTimestampFrequency UDIV"
+             underscore_name="gpu_time"
+             units="ns"
+             symbol_name="GpuTime"
+             semantic_type="duration"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="GPU Core Clocks"
+             description="The total number of GPU core clocks elapsed during the measurement."
+             data_type="uint64"
+             equation="GPU_CLOCK 0 READ"
+             underscore_name="gpu_core_clocks"
+             units="cycles"
+             symbol_name="GpuCoreClocks"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="AVG GPU Core Frequency"
+             description="Average GPU Core Frequency in the measurement."
+             data_type="uint64"
+             max_equation="$GpuMaxFrequency"
+             equation="$GpuCoreClocks 1000000000 UMUL $GpuTime UDIV"
+             underscore_name="avg_gpu_core_frequency"
+             units="hz"
+             symbol_name="AvgGpuCoreFrequency"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Tier1 Overview System Frame Batch Draw"
+             mdapi_group="GPU"
+             mdapi_hw_unit_type="gpu"
+             />
+    <counter name="TestCounter8"
+             description="HW test counter 8. Should be equal to 1."
+             data_type="uint64"
+             equation="C 7 READ"
+             underscore_name="counter8"
+             units="events"
+             symbol_name="Counter8"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter4"
+             description="HW test counter 4. Factor: 0.333"
+             data_type="uint64"
+             equation="B 4 READ"
+             underscore_name="counter4"
+             units="events"
+             symbol_name="Counter4"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter5"
+             description="HW test counter 5. Factor: 0.333"
+             data_type="uint64"
+             equation="B 5 READ"
+             underscore_name="counter5"
+             units="events"
+             symbol_name="Counter5"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter6"
+             description="HW test counter 6. Factor: 0.166"
+             data_type="uint64"
+             equation="B 6 READ"
+             underscore_name="counter6"
+             units="events"
+             symbol_name="Counter6"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter3"
+             description="HW test counter 3. Factor: 0.5"
+             data_type="uint64"
+             equation="B 3 READ"
+             underscore_name="counter3"
+             units="events"
+             symbol_name="Counter3"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter0"
+             description="HW test counter 0. Factor: 0.0"
+             data_type="uint64"
+             equation="B 0 READ"
+             underscore_name="counter0"
+             units="events"
+             symbol_name="Counter0"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter1"
+             description="HW test counter 1. Factor: 1.0"
+             data_type="uint64"
+             equation="B 1 READ"
+             underscore_name="counter1"
+             units="events"
+             symbol_name="Counter1"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <counter name="TestCounter2"
+             description="HW test counter 2. Factor: 1.0"
+             data_type="uint64"
+             equation="B 2 READ"
+             underscore_name="counter2"
+             units="events"
+             symbol_name="Counter2"
+             semantic_type="event"
+             mdapi_supported_apis=""
+             mdapi_usage_flags="Frame Batch Draw"
+             mdapi_hw_unit_type="gpu"
+             mdapi_group="GPU"
+             />
+    <register_config type="NOA">
+        <register type="NOA" address="0x00009888" value="0x11810000" />
+        <register type="NOA" address="0x00009888" value="0x07810013" />
+        <register type="NOA" address="0x00009888" value="0x1F810000" />
+        <register type="NOA" address="0x00009888" value="0x1D810000" />
+        <register type="NOA" address="0x00009888" value="0x1B930040" />
+        <register type="NOA" address="0x00009888" value="0x07E54000" />
+        <register type="NOA" address="0x00009888" value="0x1F908000" />
+        <register type="NOA" address="0x00009888" value="0x11900000" />
+        <register type="NOA" address="0x00009888" value="0x37900000" />
+        <register type="NOA" address="0x00009888" value="0x53900000" />
+        <register type="NOA" address="0x00009888" value="0x45900000" />
+        <register type="NOA" address="0x00009888" value="0x33900000" />
+    </register_config>
+    <register_config type="OA">
+        <register type="OA" address="0x00002740" value="0x00000000" />
+        <register type="OA" address="0x00002744" value="0x00800000" />
+        <register type="OA" address="0x00002714" value="0xF0800000" />
+        <register type="OA" address="0x00002710" value="0x00000000" />
+        <register type="OA" address="0x00002724" value="0xF0800000" />
+        <register type="OA" address="0x00002720" value="0x00000000" />
+        <register type="OA" address="0x00002770" value="0x00000004" />
+        <register type="OA" address="0x00002774" value="0x00000000" />
+        <register type="OA" address="0x00002778" value="0x00000003" />
+        <register type="OA" address="0x0000277C" value="0x00000000" />
+        <register type="OA" address="0x00002780" value="0x00000007" />
+        <register type="OA" address="0x00002784" value="0x00000000" />
+        <register type="OA" address="0x00002788" value="0x00100002" />
+        <register type="OA" address="0x0000278C" value="0x0000FFF7" />
+        <register type="OA" address="0x00002790" value="0x00100002" />
+        <register type="OA" address="0x00002794" value="0x0000FFCF" />
+        <register type="OA" address="0x00002798" value="0x00100082" />
+        <register type="OA" address="0x0000279C" value="0x0000FFEF" />
+        <register type="OA" address="0x000027A0" value="0x001000C2" />
+        <register type="OA" address="0x000027A4" value="0x0000FFE7" />
+        <register type="OA" address="0x000027A8" value="0x00100001" />
+        <register type="OA" address="0x000027AC" value="0x0000FFE7" />
+    </register_config>
+  </set>
+
+</metrics>
-- 
2.7.4