[AMDGPU] Fixed occupancy calculation for gfx10

Differential Revision: https://reviews.llvm.org/D65010

llvm-svn: 366616

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
index 1eb9b83..716c387 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.cpp
@@ -591,25 +591,12 @@ unsigned GCNSubtarget::getOccupancyWithNumSGPRs(unsigned SGPRs) const {
 }
 
 unsigned GCNSubtarget::getOccupancyWithNumVGPRs(unsigned VGPRs) const {
-  if (VGPRs <= 24)
-    return 10;
-  if (VGPRs <= 28)
-    return 9;
-  if (VGPRs <= 32)
-    return 8;
-  if (VGPRs <= 36)
-    return 7;
-  if (VGPRs <= 40)
-    return 6;
-  if (VGPRs <= 48)
-    return 5;
-  if (VGPRs <= 64)
-    return 4;
-  if (VGPRs <= 84)
-    return 3;
-  if (VGPRs <= 128)
-    return 2;
-  return 1;
+  unsigned MaxWaves = getMaxWavesPerEU();
+  unsigned Granule = getVGPRAllocGranule();
+  if (VGPRs < Granule)
+    return MaxWaves;
+  unsigned RoundedRegs = ((VGPRs + Granule - 1) / Granule) * Granule;
+  return std::min(getTotalNumVGPRs() / RoundedRegs, MaxWaves);
 }
 
 unsigned GCNSubtarget::getReservedNumSGPRs(const MachineFunction &MF) const {

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 78c3b82..1f30d76 100644 (file)
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -884,7 +884,7 @@ public:
   /// \returns Maximum number of waves per execution unit supported by the
   /// subtarget without any kind of limitation.
   unsigned getMaxWavesPerEU() const {
-    return AMDGPU::IsaInfo::getMaxWavesPerEU();
+    return AMDGPU::IsaInfo::getMaxWavesPerEU(this);
   }
 
   /// \returns Number of waves per work group supported by the subtarget and

diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
index e90f40e..ef50d37 100644 (file)
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -241,7 +241,7 @@ unsigned getMaxWorkGroupsPerCU(const MCSubtargetInfo *STI,
 }
 
 unsigned getMaxWavesPerCU(const MCSubtargetInfo *STI) {
-  return getMaxWavesPerEU() * getEUsPerCU(STI);
+  return getMaxWavesPerEU(STI) * getEUsPerCU(STI);
 }
 
 unsigned getMaxWavesPerCU(const MCSubtargetInfo *STI,
@@ -253,9 +253,11 @@ unsigned getMinWavesPerEU(const MCSubtargetInfo *STI) {
   return 1;
 }
 
-unsigned getMaxWavesPerEU() {
+unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI) {
   // FIXME: Need to take scratch memory into account.
-  return 10;
+  if (!isGFX10(*STI))
+    return 10;
+  return 20;
 }
 
 unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI,
@@ -317,7 +319,7 @@ unsigned getMinNumSGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU) {
   if (Version.Major >= 10)
     return 0;
 
-  if (WavesPerEU >= getMaxWavesPerEU())
+  if (WavesPerEU >= getMaxWavesPerEU(STI))
     return 0;
 
   unsigned MinNumSGPRs = getTotalNumSGPRs(STI) / (WavesPerEU + 1);
@@ -394,17 +396,19 @@ unsigned getVGPREncodingGranule(const MCSubtargetInfo *STI,
 }
 
 unsigned getTotalNumVGPRs(const MCSubtargetInfo *STI) {
-  return 256;
+  if (!isGFX10(*STI))
+    return 256;
+  return STI->getFeatureBits().test(FeatureWavefrontSize32) ? 1024 : 512;
 }
 
 unsigned getAddressableNumVGPRs(const MCSubtargetInfo *STI) {
-  return getTotalNumVGPRs(STI);
+  return 256;
 }
 
 unsigned getMinNumVGPRs(const MCSubtargetInfo *STI, unsigned WavesPerEU) {
   assert(WavesPerEU != 0);
 
-  if (WavesPerEU >= getMaxWavesPerEU())
+  if (WavesPerEU >= getMaxWavesPerEU(STI))
     return 0;
   unsigned MinNumVGPRs =
       alignDown(getTotalNumVGPRs(STI) / (WavesPerEU + 1),

diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
index 209ef7e..590df20 100644 (file)
--- a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
+++ b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -94,7 +94,7 @@ unsigned getMinWavesPerEU(const MCSubtargetInfo *STI);
 
 /// \returns Maximum number of waves per execution unit for given subtarget \p
 /// STI without any kind of limitation.
-unsigned getMaxWavesPerEU();
+unsigned getMaxWavesPerEU(const MCSubtargetInfo *STI);
 
 /// \returns Maximum number of waves per execution unit for given subtarget \p
 /// STI and limited by given \p FlatWorkGroupSize.

diff --git a/llvm/test/CodeGen/AMDGPU/hsa-metadata-kernel-code-props-v3.ll b/llvm/test/CodeGen/AMDGPU/hsa-metadata-kernel-code-props-v3.ll
index 0eed325..2d2f9ce 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/hsa-metadata-kernel-code-props-v3.ll
+++ b/llvm/test/CodeGen/AMDGPU/hsa-metadata-kernel-code-props-v3.ll
@@ -74,7 +74,10 @@ entry:
 
 ; CHECK:   .name:       num_spilled_vgprs
 ; CHECK:   .symbol:     num_spilled_vgprs.kd
-; CHECK:   .vgpr_spill_count: 14
+; GFX700:   .vgpr_spill_count: 14
+; GFX803:   .vgpr_spill_count: 14
+; GFX900:   .vgpr_spill_count: 14
+; GFX1010:  .vgpr_spill_count: 0
 define amdgpu_kernel void @num_spilled_vgprs() #1 {
   %val0 = load volatile float, float addrspace(1)* @var
   %val1 = load volatile float, float addrspace(1)* @var

diff --git a/llvm/test/CodeGen/AMDGPU/idot8s.ll b/llvm/test/CodeGen/AMDGPU/idot8s.ll
index 38cd115..3667dfc 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/idot8s.ll
+++ b/llvm/test/CodeGen/AMDGPU/idot8s.ll
@@ -2356,9 +2356,9 @@ define amdgpu_kernel void @idot8_acc8_vecMul(<8 x i4> addrspace(1)* %src1,
 ; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v13, 12, s6
 ; GFX10-DL-NEXT:    v_mul_lo_u16_e64 v7, v22, v11
 ; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v15, 12, s0
-; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v20, 12, s8
+; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v31, 12, s8
 ; GFX10-DL-NEXT:    v_mul_lo_u16_e64 v23, v23, v10
-; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v21, 12, s1
+; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v27, 12, s1
 ; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v16, 12, s5
 ; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v17, 12, s9
 ; GFX10-DL-NEXT:    v_lshlrev_b16_e64 v19, 12, s7
@@ -2368,8 +2368,8 @@ define amdgpu_kernel void @idot8_acc8_vecMul(<8 x i4> addrspace(1)* %src1,
 ; GFX10-DL-NEXT:    v_and_b32_e32 v12, v16, v2
 ; GFX10-DL-NEXT:    v_and_b32_e32 v13, v17, v2
 ; GFX10-DL-NEXT:    v_and_b32_e32 v15, v19, v2
-; GFX10-DL-NEXT:    v_and_b32_e32 v10, v21, v2
-; GFX10-DL-NEXT:    v_and_b32_e32 v14, v20, v2
+; GFX10-DL-NEXT:    v_and_b32_e32 v10, v27, v2
+; GFX10-DL-NEXT:    v_and_b32_e32 v14, v31, v2
 ; GFX10-DL-NEXT:    v_and_b32_sdwa v6, v23, v2 dst_sel:BYTE_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
 ; GFX10-DL-NEXT:    v_and_b32_sdwa v7, v7, s2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX10-DL-NEXT:    v_and_b32_sdwa v4, v4, s2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD

diff --git a/llvm/test/CodeGen/AMDGPU/nsa-reassign.ll b/llvm/test/CodeGen/AMDGPU/nsa-reassign.ll
index a668a19..011aef8 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/nsa-reassign.ll
+++ b/llvm/test/CodeGen/AMDGPU/nsa-reassign.ll
@@ -21,8 +21,8 @@ main_body:
 }
 
 ; GCN-LABEL: {{^}}sample_contig_nsa_10vgprs:
-; GCN-DAG: image_sample_c_l v{{[0-9]+}}, [{{v[0-9]+, v[0-9]+, v[0-9]+, v[0-9]+, v[0-9]+}}],
-; GCN-DAG: image_sample v{{[0-9]+}}, [{{v[0-9]+, v[0-9]+, v[0-9]+}}],
+; GCN-DAG: image_sample_c_l v{{[0-9]+}}, v[{{[0-9:]+}}],
+; GCN-DAG: image_sample v{{[0-9]+}}, v[{{[0-9:]+}}],
 define amdgpu_ps <2 x float> @sample_contig_nsa_10vgprs(<8 x i32> inreg %rsrc, <4 x i32> inreg %samp, float %zcompare, float %s1, float %t1, float %r1, float %lod, float %r2, float %s2, float %t2) #0 {
 main_body:
   %zcompare.1 = fadd float %zcompare, 1.0

diff --git a/llvm/test/CodeGen/AMDGPU/regbank-reassign.mir b/llvm/test/CodeGen/AMDGPU/regbank-reassign.mir
index eb7b706..450439f 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/regbank-reassign.mir
+++ b/llvm/test/CodeGen/AMDGPU/regbank-reassign.mir
@@ -217,6 +217,12 @@ registers:
   - { id: 7, class: vreg_128, preferred-register: '$vgpr12_vgpr13_vgpr14_vgpr15' }
   - { id: 8, class: vreg_128, preferred-register: '$vgpr16_vgpr17_vgpr18_vgpr19' }
   - { id: 9, class: vreg_128, preferred-register: '$vgpr20_vgpr21_vgpr22_vgpr23' }
+  - { id: 10, class: vreg_128, preferred-register: '$vgpr24_vgpr25_vgpr26_vgpr27' }
+  - { id: 11, class: vreg_128, preferred-register: '$vgpr28_vgpr29_vgpr30_vgpr31' }
+  - { id: 12, class: vreg_128, preferred-register: '$vgpr32_vgpr33_vgpr34_vgpr35' }
+  - { id: 13, class: vreg_128, preferred-register: '$vgpr36_vgpr37_vgpr38_vgpr39' }
+  - { id: 14, class: vreg_128, preferred-register: '$vgpr40_vgpr41_vgpr42_vgpr43' }
+  - { id: 15, class: vreg_128, preferred-register: '$vgpr44_vgpr45_vgpr46_vgpr47' }
 body: |
   bb.0:
     %0 = IMPLICIT_DEF
@@ -228,6 +234,12 @@ body: |
     %7 = IMPLICIT_DEF
     %8 = IMPLICIT_DEF
     %9 = IMPLICIT_DEF
+    %10 = IMPLICIT_DEF
+    %11 = IMPLICIT_DEF
+    %12 = IMPLICIT_DEF
+    %13 = IMPLICIT_DEF
+    %14 = IMPLICIT_DEF
+    %15 = IMPLICIT_DEF
     %2 = V_AND_B32_e32 %1, %0, implicit $exec
     GLOBAL_STORE_DWORD %3, %0, 0, 0, 0, 0, implicit $exec
     GLOBAL_STORE_DWORD %3, %1, 0, 0, 0, 0, implicit $exec

diff --git a/llvm/test/CodeGen/AMDGPU/wave32.ll b/llvm/test/CodeGen/AMDGPU/wave32.ll
index a71ca5d..f4715a2 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/wave32.ll
+++ b/llvm/test/CodeGen/AMDGPU/wave32.ll
@@ -920,7 +920,7 @@ main_body:
 
 ; GCN-LABEL: {{^}}test_vgprblocks_w64_attr:
 ; Test that the wave size can be overridden in function attributes and that the block size is correct as a result
-; GFX10DEFWAVE: ; VGPRBlocks: 2
+; GFX10DEFWAVE: ; VGPRBlocks: 11
 define amdgpu_gs float @test_vgprblocks_w64_attr(float %a, float %b, float %c, float %d, float %e,
                                         float %f, float %g, float %h, float %i, float %j, float %k, float %l) #4 {
 main_body: