zink: add VK_BUFFER_USAGE_INDEX_BUFFER_BIT to vertex buffer creation

these can be mixed vertex/index buffers, and there's not really any way
to know in advance

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7132>

zink: handle null attachment for ARB_texture_buffer_object samplers

according to spec:

If no buffer object is bound to the buffer texture,
the results of the texel access are undefined.

so this is technically allowed

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7132>

zink: assert valid format in zink_create_sampler_view()

catch issues higher up the stack here

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7132>

zink: ensure resource tracking for sampler buffers in render batches

this fixes some synchronization issues when using e.g., buffer textures
that change their data repeatedly

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7132>

zink: implement ARB_texture_buffer_object

the pipe cap for this was enabled for some reason, but the actual functionality
was never implemented

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7132>

zink: increase descriptor pool sizes for other descriptor types we'll be using

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7132>

glsl: more accurately handle swizzle in 64bit varying split with no left value

as implied in the surrounding code, left_components can be 0 here, in which
case creating a left swizzle is unnecessary (and triggers an assert)

this moves a failing assert farther down the stack to a more useful location
when trying to pack e.g., struct[3] { dvec3; float; }

ref spec@arb_gpu_shader_fp64@execution@inout@vs-out-fs-in-s1-s2@3-dvec2-float

Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7134>

aco: don't use v_pack_b32_f16 if 16-bit input denormals are flushed

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7111>

aco: propagate literals into sub-dword pseudo instructions on GFX9+

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7111>

aco: implement 16-bit literals

We can copy any value into a 16-bit subregister with a 3 dword
v_pack_b32_f16 on GFX10 or a v_and_b32+v_or_b32 on GFX9.

Because the generated code can depend on the register assignment and to
improve constant propagation, Builder::copy creates a p_create_vector in
the case of sub-dword literals.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7111>

panfrost: Add blend shader support to bifrost

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

panfrost: Add missing tile-buffer formats to the format enum

Some tile-buffer formats are missing, add them.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Special-case load_input for blend shaders

Blend shaders are passed blend inputs through r0-r3. Let's emit a MOV
from those register when we see a load_input intrinsic.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Reserve r0-r3 in blend shaders

Blend shaders are passed the source color through r0-r3. Let's avoid
allocating those. The is definitely not the right solution but is good
enough for now.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Special-case BLEND instruction emission for blend shaders

Blend shaders shouldn't use the blend descriptors stored in the FAU RAM
since this is what triggered the blend shader call in the first place.
The descriptor is instead extracted from the compiler inputs and passed
as a constant to the blend instruction.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Collect return addresses of blend calls

We will need that for blend shaders so they can be passed a return
address and jump back to the fragment shader when they're done.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Add load_output support

This is mapped to the LD_TILE instruction. Note that multi-sample RTs
are not supported yet.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

panfrost: Flag blend shader function as an entry point

Some lowering functions used by bifrost are searching for an entry point.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

panfrost: Scalarize nir_load_blend_const_color_rgba

Bifrost is a scalar architecture, which means we can't load all
components of the blend constant at once. We could add a lowering pass
to scalarize nir_load_blend_const_color_rgba, but it's easier to handle
that at when lowering the blend equations.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

panfrost: Add a "Bifrost Internal Blend" descriptor

This descriptor can be passed directly as a constant to the bifrost
BLEND instruction and we'll need to pass this information to blend
shaders. Let's extract the "Bifrost Internal Blend" descriptor from the
"Bifrost Blend Overlay" definition.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Support indirect jumps

We need that for blend shaders which are passed the return address
through r48.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Add support for load_blend_const_color_{r,g,b,a}_float

Needed for blend shaders.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Rework blend descriptor access handling

The current logic assumes blend descriptors are always retrieved from
the blend descriptor slots present in the FAU RAM, but this assumption
no longer stands when we add blend shaders to the mix. In that case we
need to use an 'opaque blend' whose descriptor is passed through
embedded constants.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Get rid of the regs argument in bi_assign_fau_idx()

Regs are already part of the bundle struct, let's just pass a pointer
to this bundle object instead of passing both the bundle and regs.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Use canonical name for FAU RAM sources

The uniform_constant field and BIFROST_SRC_CONST_{LO,HI} definitions
seem to imply that those only deal with embedded constants. Let's
rename them to reflect the fact that they actually encode accesses to
the Fast-Access-Uniform RAM.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

pan/bi: Copy blend shader info from compile_inputs

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

panfrost: Extend compile_inputs to pass a blend descriptor

This is needed for BLEND instructions used from a blend shader so we can
store the result of the shader-based blending back to the tile buffer.
We let the gallium driver build this blend descriptor for us in order
to keep the compiler cmdstream-agnostic.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

panfrost: Fix fixed-function blend on bifrost

The conversion from a 32b float to a 16b fixed-point number was wrong.

Fixes: 8389976b7c09 ("panfrost: XML-ify the blend descriptors")
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reviewed-by: Alyssa Rosenzweig <alyssa.rosenzweig@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7151>

v3d/compiler: implement load interpolated input intrinsics

We will lower GLSL interpolateAt functions to these.

Reviewed-by: Alejandro Piñeiro <apinheiro@igalia.com>
Acked-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7155>

broadcom/compiler: track partially interpolated fragment inputs

We will need these to implement GLSL's interpolateAt*() functions where
we are required to perform interpolation in the shader at arbitrary
offsets.

Acked-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7155>

iris: Fix doubling of shared local memory (SLM) sizes.

Commit 67ee9c5f5537fe85357556a4322a07253d13a697 added support for
using the `pipe_compute_state::req_local_mem` field, because Clover
can have a run-time specified size that isn't baked into the shaders.

However, it started adding the static size from the shader to the
dynamic state-supplied size.  The Mesa state tracker fills out
req_local_mem to prog->Base.info.cs.shared_size, which is exactly
what we fill out prog_data->total_shared to be.  Effectively, this
meant that we double-counted the same SLM requirements, doubling
our space requirements.

Fixes a 10% performance regression in Synmark2's OglCSDof test.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7152>

intel/compiler, anv: Delete cs_prog_data->slm_size

cs_prog_data->slm_size is basically redundant with
prog_data->total_shared, which is the field that we actually use for
controlling the shared local memory size in all drivers.  We were
still using it in one place for VK_EXT_pipeline_executable_properties,
but we should just fix that and delete the field.

Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7152>

broadcom/compiler: use nir io semantics

This allows to clean up some code.

Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6721>

nir/lower_io_to_scalar: update io semantics on per-component inst

When we replace the original instruction with per-channel operations,
the new instruction should inherint the semantics of the original
instruction.

Reviewed-by: Eric Anholt <eric@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6721>

broadcom/compiler: support varyings with struct types

This adds support for using structs as outputs from vertex shaders and
inputs to fragment shaders.

Reviewed-by: Iago Toral Quiroga <itoral@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6721>

docs/release-calendar: plan 20.3 release

Signed-off-by: Eric Engestrom <eric@engestrom.ch>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6616>

intel/fs: Allow constant-propagation into SAMPLEINFO and IMAGE_SIZE

Without this, we end up with indirect sampler messages all the time
because we don't propagate the texture/image BTI.  This makes debugging
shaders with imageSize or textureSamples in them a pain.

Shader-db results on Ice Lake:

    total instructions in shared programs: 19720612 -> 19720564 (<.01%)
    instructions in affected programs: 4998 -> 4950 (-0.96%)
    helped: 12
    HURT: 0

All affected shaders were compute shaders in Deus Ex: Mankind Divided.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6794>

docs: update calendar and link releases notes for 20.1.10

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7147>

docs: add release notes for 20.1.10

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7147>

isl: Allow CCS for 8bpp surfaces with 3+ miplevels

I can't find a restriction for enabling CCS on these surfaces in recent
versions of the Bspec. Since I didn't cite my source, I'm not even sure
such a restriction existed in the first place.

Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7085>

iris: Add fast-clear restriction for 8bpp surfaces

For 8bpp surfaces on TGL, prevent LOD1+ from being fast-cleared. This
will be relevant once ISL starts allowing CCS for 8bpp surfaces with
more than 2 miplevels. I verified the problem behind this restriction
with a modified version of the fbo-clearmipmap piglit test.

Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7085>

docs: update calendar and link releases notes for 20.2.1

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7141>

docs: add SHA256 sums for 20.2.1

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7141>

docs: add release notes for 20.2.1

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7141>

radv: fix optimizing needed states if some are marked as dynamic

From the Vulkan spec 1.2.157:

    "VK_DYNAMIC_STATE_STENCIL_TEST_ENABLE_EXT specifies that the
     stencilTestEnable state in VkPipelineDepthStencilStateCreateInfo
     will be ignored and must be set dynamically with
     vkCmdSetStencilTestEnableEXT before any draw call."

So, stencilTestEnable should be ignored if dynamic. While we are
at it, fix depthBoundsTestEnable too.

Cc: 20.2
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3633
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7112>

docs: Document how to replicate a CI build locally.

Who hasn't needed to do this at some point?  Turns out it's not too hard
to do, and was useful for me in iterating on the Android build.

Acked-by: Michel Dänzer <mdaenzer@redhat.com>
Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6700>

ci/android: Switch to using the Android NDK.

To support Android drivers, we're going to want to be tracking that Mesa's
build succeeds on a real android toolchain.  This still uses the android
stubs since these libs aren't in the NDK.

Note that I had to drop the Intel and AMD drivers currently: we don't have
LLVM cross-compiled for Android in this container, and I'm honestly hoping
ACO saves us from that.  Intel has dependencies on libexpat, which AOSP
really doesn't want to bring in, and it looks to me like those dependencies
could be optional.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6700>

symbols-check: Add __cxa_guard_* to the list of approved symbols.

These are introduced by the compiler during static local initialization in
c++ for thread safety.  This seems to end up being public in the driver
with --static-libc++ on android.

Reviewed-by: <ajax@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6700>

glsl/tests: Make the tests skip on Android binary execution failures.

We don't have a suitable exe wrapper for running them, and the missing
linker is throwing return code 255 instead of an ENOEXEC.  Catch it and
return skip from the tests.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6700>

meson: Drop adding -Wl,--gc-sections to project c/cpp arguments.

We already have the targets we care about doing this using
ld_args_gc_sections, and by adding it to project arguments we caused
warnings spam in the android clang build about the compile stage not using
the argument.

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6700>

aco/isel: Remove now unused VS-related code from create_null_export

Also replaced a hardcoded constant with the appropriate register macro.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7102>

aco/isel: Remove some dead code

exported_pos was always initialized to true (due to the is_pos argument
of the first export_vs_varying call being true), so none of this code has
any effect.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7102>

aco/isel: Always export position data from VS/NGG

AMD ISA docs explicitly require this for VS, and this likely extends to
NGG too.

Cc: mesa-stable
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3615
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7102>

aco: use VOP2 for v_cvt_pkrtz_f16_f32 if possible

This patch also does a slight rework of export_fs_mrt_color()
to avoid setting of enabled channels which are not used.

Totals from 52404 (38.38% of 136546) affected shaders (NAVI):
SGPRs: 3097443 -> 3097435 (-0.00%)
CodeSize: 189151600 -> 188546200 (-0.32%)
Instrs: 36445061 -> 36445104 (+0.00%); split: -0.00%, +0.00%
Cycles: 1739388020 -> 1739388192 (+0.00%); split: -0.00%, +0.00%
VMEM: 21071501 -> 21071665 (+0.00%); split: +0.00%, -0.00%
SMEM: 3470983 -> 3470982 (-0.00%); split: +0.00%, -0.00%
PreSGPRs: 2058965 -> 2058962 (-0.00%)
PreVGPRs: 1860294 -> 1860295 (+0.00%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

aco: use VOP2 version of v_cvt_pkrtz_f16_f32 on GFX_6_7_10

Totals from 767 (0.56% of 136546) affected shaders (NAVI):
CodeSize: 2862208 -> 2850036 (-0.43%)
Instrs: 561572 -> 561574 (+0.00%)
Cycles: 6455420 -> 6455428 (+0.00%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

radv,aco: lower_pack_half_2x16

This patch also optimizes pack_half_2x16(a, 0.0).

Totals from 1949 (1.43% of 136546) affected shaders (RAVEN):
SGPRs: 83376 -> 83336 (-0.05%)
CodeSize: 3532144 -> 3512352 (-0.56%)
Instrs: 660746 -> 660682 (-0.01%); split: -0.01%, +0.00%
Cycles: 6780716 -> 6780472 (-0.00%); split: -0.00%, +0.00%
VMEM: 990886 -> 990883 (-0.00%); split: +0.00%, -0.00%
SMEM: 150506 -> 150538 (+0.02%); split: +0.05%, -0.03%
SClause: 30595 -> 30594 (-0.00%); split: -0.01%, +0.00%
Copies: 40801 -> 40729 (-0.18%)
PreSGPRs: 52335 -> 52341 (+0.01%); split: -0.03%, +0.04%
PreVGPRs: 45104 -> 45097 (-0.02%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

aco: use v_cvt_pkrtz_f16_f32 for pack_half_2x16

Apparently, we forgot to remove some debug code.
This patch also fixes the round mode check to consider
the destination bit width.

Totals from 2218 (1.62% of 136546) affected shaders (RAVEN):
SGPRs: 100848 -> 100280 (-0.56%)
VGPRs: 68536 -> 66044 (-3.64%); split: -3.68%, +0.05%
CodeSize: 4882296 -> 4837220 (-0.92%); split: -0.94%, +0.01%
MaxWaves: 18990 -> 19019 (+0.15%); split: +0.19%, -0.04%
Instrs: 938150 -> 930388 (-0.83%); split: -0.83%, +0.00%
Cycles: 8699824 -> 8667648 (-0.37%); split: -0.38%, +0.01%
VMEM: 1144502 -> 1059680 (-7.41%); split: +0.06%, -7.48%
SMEM: 170076 -> 167999 (-1.22%); split: +0.22%, -1.44%
VClause: 18428 -> 18422 (-0.03%)
SClause: 41375 -> 41353 (-0.05%); split: -0.06%, +0.00%
Copies: 60008 -> 60054 (+0.08%); split: -0.31%, +0.39%
PreVGPRs: 56163 -> 56142 (-0.04%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

aco: add validation rules for p_split_vector

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

aco: use p_split_vector for nir_op_unpack_half_*

This enables the use of SDWA if possible

Totals from 9933 (7.27% of 136546) affected shaders (RAVEN):
VGPRs: 731764 -> 731772 (+0.00%); split: -0.00%, +0.00%
CodeSize: 90944852 -> 90671472 (-0.30%); split: -0.30%, +0.00%
Instrs: 17881885 -> 17867831 (-0.08%); split: -0.08%, +0.00%
Cycles: 1597904072 -> 1597771260 (-0.01%); split: -0.01%, +0.00%
VMEM: 1702328 -> 1697383 (-0.29%); split: +0.13%, -0.42%
SMEM: 659583 -> 659049 (-0.08%); split: +0.01%, -0.09%
VClause: 318024 -> 318025 (+0.00%); split: -0.00%, +0.00%
SClause: 631670 -> 631707 (+0.01%); split: -0.01%, +0.01%
Copies: 1504107 -> 1504626 (+0.03%); split: -0.01%, +0.04%
PreVGPRs: 683153 -> 683180 (+0.00%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

nir/opt_algebraic: optimize unpack_half_2x16_split_x(ushr, a, 16)

Same as extract_u16(a, 1)

Totals from 2021 (1.48% of 136546) affected shaders (RAVEN):
VGPRs: 129516 -> 129524 (+0.01%); split: -0.00%, +0.01%
CodeSize: 12485704 -> 12486600 (+0.01%); split: -0.00%, +0.01%
Instrs: 2435041 -> 2434999 (-0.00%); split: -0.00%, +0.00%
Cycles: 20952552 -> 20952624 (+0.00%); split: -0.00%, +0.00%
VMEM: 374492 -> 374212 (-0.07%); split: +0.01%, -0.08%
SMEM: 123309 -> 123291 (-0.01%); split: +0.00%, -0.02%
VClause: 64156 -> 64164 (+0.01%)
Copies: 191620 -> 191616 (-0.00%); split: -0.03%, +0.03%

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

aco: use p_create_vector for nir_op_pack_half_2x16

This enables the use of SDWA if possible

Totals from 2218 (1.62% of 136546) affected shaders (RAVEN):
VGPRs: 68508 -> 68516 (+0.01%)
CodeSize: 4897024 -> 4881068 (-0.33%); split: -0.33%, +0.00%
MaxWaves: 18992 -> 18990 (-0.01%)
Instrs: 946942 -> 939161 (-0.82%); split: -0.82%, +0.00%
Cycles: 8737668 -> 8705704 (-0.37%); split: -0.37%, +0.00%
VMEM: 1155362 -> 1145245 (-0.88%); split: +0.00%, -0.88%
SMEM: 170435 -> 170165 (-0.16%); split: +0.01%, -0.16%
VClause: 18426 -> 18425 (-0.01%)
SClause: 41376 -> 41375 (-0.00%)
Copies: 59813 -> 59787 (-0.04%); split: -0.15%, +0.10%
PreVGPRs: 56126 -> 56136 (+0.02%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

aco: expand create_vector more carefully w.r.t. subdword operands

No pipelinedb changes.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

aco: propagate SGPRs into VOP1 instructions early.

This helps DCE. We should reconsider our optimization order
or maybe do the dead code analysis twice

Totals from 106 (0.08% of 136546) affected shaders (RAVEN):
SGPRs: 7184 -> 7152 (-0.45%)
CodeSize: 736912 -> 736052 (-0.12%)
Instrs: 145739 -> 145509 (-0.16%)
Cycles: 2085344 -> 2084268 (-0.05%)
VMEM: 14819 -> 14807 (-0.08%)
SMEM: 7109 -> 7100 (-0.13%); split: +0.04%, -0.17%
SClause: 5383 -> 5385 (+0.04%)
Copies: 13290 -> 13189 (-0.76%)
PreSGPRs: 5265 -> 5221 (-0.84%)

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6777>

zink: unify code for emitting named uint-based variable instructions

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7130>

aco: adjust an assertion about the wavesize in emit_gfx10_wave64_bpermute()

This gets rids of one more use of radv_shader_info.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7061>

aco: compute the CS workgroup size from the shader NIR info

cs.block_size is copied from cs.local_size during the shader info pass.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7061>

radv: move compiler statistics to ACO

They are really specific to ACO.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7061>

aco: remove unused radv_shader.h includes

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7061>

aco: remove useless occurences of radv_nir_compiler_options

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7061>

aco: remove stub lower_wqm() prototype

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7061>

zink: export PIPE_CAP_MAX*_VARYINGS values

this is separate from PIPE_SHADER_CAP_MAX_OUTPUTS

fixes mesa/mesa#3105

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7110>

zink: add feature-documentation

This adds some documentation for the current feature-set in Zink,
explaining what extensions are currently needed for what functionality.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7116>

zink: redo slot mapping again for the last time really I mean it

now that shader compiling is happening all at once, we can store the slot
map on zink_gfx_program directly and reserve it dynamically in order to
use up only the slots that are actually being used across all shader stages

Reviewed-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7100>

zink: don't leak sampler view textures

by adding a batch reference for these textures during draw, we can successfully
destroy the resources without crashing

Reviewed-by: Erik Faye-Lun <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6924>

zink: explicitly flag fb attachments as being written to in render passes

we need to ensure that we're accurately setting this hint in order to avoid
synchronization issues when determining whether we can read from the buffer

Reviewed-by: Erik Faye-Lun <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6924>

zink: add more explicit fencing for transfer maps

we're using our (primitive) buffer r/w tracking here to ensure that our
src buffers are synchronized before we do any kind of read operation on them

this is pretty slow in some cases, but it fixes a bunch of piglit tests

Reviewed-by: Erik Faye-Lun <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6924>

zink: optimize transfer_map for resources with pending reads/writes

we don't need to stall here if we know that we're not about to have any io
conflicts in the buffer

Reviewed-by: Erik Faye-Lun <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6924>

zink: add a mechanism to track current resource usage in batches

this is really primitive, but it at least gives an idea of whether a
resource has been submitted for writing in a pending batch

Reviewed-by: Erik Faye-Lun <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6924>

radv: fix ignoring the vertex attribute stride if set as dynamic

The vertex attribute stride should be ignored, so make sure it's
initialized to zero if dynamic to avoid computing a wrong offset.

The fact that each element of pStrides must be greater than or equal
to the maximum extent of all vertex input attributes fetched saves us
one user SGPR for the dynamic stride.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/3627
Cc: 20.2
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7101>

ac,amd/llvm,radv: Initialize structs with {0}

Necessary to compile with MSVC.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7123>

radv/aco: disable NGG GS support because it randomly hangs the GPU

Disable ACO NGG GS until the random GPU hangs are fixed
(one CTS run == one GPU hang here). No hangs so far after
5 full CTS runs with this disabled.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Acked-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7108>

nir/opt_uniform_atomics: remove useless returns

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7117>

radv: Only close local_fd when valid

Necessary when drm_device is bypassed.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7119>

util: Hide timespec_passed on Windows

Windows doesn't have clockid_t.

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7119>

radv: Increased const usage

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7119>

amd/addrlib: Fix warning list for msvc

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7119>

intel/fs: Rework scratch handling on Gen9+

The current scratch mechanism uses an MRF hack where we reserve a few
GRF registers to treat like the MRF and we collect the data into that
MRF region before doing a scratch write.  We also use that region for
the header for scratch reads.

This commit changes things and gets rid of the MRF hack.  Instead, we
reserve a single register (which RA is free to pick) for the scratch
header and uses split sends for scratch writes to avoid having to do
the copy.  This should provide RA with more freedom in the presence of
spilling as well as avoid some unnecessary data moves.  In future, the
new GEN9_SCRATCH_HEADER opcode gives us a place where we can do our own
per-thread scratch base address calculations rather than depending on
the scratch base address that gets pushed into g0.  Having an opcode for
this lets us do it once at the top of the shader rather than repeating
it at every read/write.

One other noticeable difference is the use of SHADER_OPCODE_SEND.  We
can get away with this thanks to the fact that we're now using a set to
track which instructions are generated by spills and don't rely on the
opcodes to find spill/fill instructions.  This allows us to avoid adding
more virtual opcodes and let the normal code paths handle things like
scoreboard dependencies between header setup and the SEND.  It also
means that post-RA scheduling may be able to space out the header setup
MOV and the SEND for better latency hiding.

Shader-db results on Skylake:

    total spills in shared programs: 12137 -> 10604 (-12.63%)
    spills in affected programs: 6685 -> 5152 (-22.93%)
    helped: 274
    HURT: 2

    total fills in shared programs: 13065 -> 11515 (-11.86%)
    fills in affected programs: 9007 -> 7457 (-17.21%)
    helped: 275
    HURT: 1

Shader-db results on Ice Lake:

    total spills in shared programs: 12482 -> 10953 (-12.25%)
    spills in affected programs: 6586 -> 5057 (-23.22%)
    helped: 275
    HURT: 0

    total fills in shared programs: 12819 -> 11234 (-12.36%)
    fills in affected programs: 7867 -> 6282 (-20.15%)
    helped: 274
    HURT: 0

Shader-db results on Tigerlake:

    total spills in shared programs: 11689 -> 10233 (-12.46%)
    spills in affected programs: 4740 -> 3284 (-30.72%)
    helped: 259
    HURT: 0

    total fills in shared programs: 10840 -> 9443 (-12.89%)
    fills in affected programs: 6244 -> 4847 (-22.37%)
    helped: 259
    HURT: 0

Fossil-db results on Ice Lake:

    Spills in all programs: 245249 -> 201633 (-17.8%)
    Fills in all programs: 366066 -> 314368 (-14.1%)

More practically, this seems to give about a 0.5-1% perf boost in
Witcher 3 (DXVK) and Shadow of the Tomb Raider (Vulkan native).

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/fs/ra: Use a set to track added spill/fill instructions

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/fs/ra: Sanity-check our IP counts

Starting with e99081e76d4a, we don't re-construct liveness information
every time we spill a register.  Instead, we're very careful to track
which instructions are spill instructions and not contribute those to
the IP count so that we can continue to use the old liveness information
even though instructions have been added.  This commit adds an assert
that sanity-checks that we count the same number of instructions as our
liveness information is based on.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/fs/ra: Store the last non-spill VGRF node

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/fs/ra: Refactor handling of Gen7 scratch reads

The attempt at de-duplication with the gen7_read Boolean wasn't actually
saving us anything.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/fs/ra: Increment spill_offset as part of the emit_spill loop

This makes it consistent with our handling of src.offset and with our
handling of spill_offset in emit_unspill.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/fs: Add a SCRATCH_HEADER opcode

This opcode is responsible for setting up the buffer base address and
per-thread scratch space fields of a scratch message header.  For the
most part, it's a copy of g0 but some messages need us to zero out g0.2
and the bottom bits of g0.5.

This may actually fix a bug when nir_load/store_scratch is used.  The
docs say that the DWORD scattered messages respect the per-thread
scratch size specified in gN.3[3:0] in the message header but we've been
leaving it zero.  This may mean that we've been ignoring any scratch
reads/writes from a load/store_scratch intrinsic above the 1KB mark.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/fs: Copy the PTSS from g0 for scratch reads/writes

In theory, this fixes a bug where we were dropping the PTSS bound on the
floor.  The hardware docs claim that the A32 DWORD and BYTE scattered
read/write messages do a PTSS bounds check.   However, in practice, it
seems that the hardware ignores the bounds check so this doesn't
actually matter.  I verified this with the following couple of piglit
tests:

    https://gitlab.freedesktop.org/mesa/piglit/-/merge_requests/399

In practice, this prevents the next commit from making a subtle
behavioral change.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

intel/batch_decoder: Don't clame vec4 vs/gs/tcs shaders on Gen11+

Because we hard-coded the default to vec4, any platform where it doesn't
have a "Dispatch Mode" field gets vec4 by default.  This includes Gen11+
where vec4 is no longer a thing.  Change the default so it works on
newer hardware.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/7084>

v3dv/device: Support loader interface version 3.

Port of 1e41d7f7b0855934744fe578ba4eae9209ee69f7:
"anv: Support loader interface version 3 (patch v2)"

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: fix buffer copies to compressed images on the blit path

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: drop a couple of obsolete comments

We only expose a coherent memory heap, so invalidation and flushing
are always no-ops for us.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: limit blit framebuffer dimensions to max coordinates

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>

v3dv: generate proper UUIDs for device and driver

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/6766>