panfrost: Cache number of users of a resource

This can be tracked efficiently with atomics, and reduces the places we
use the rsrc->track.users bitmap which has concurrency issues.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12528>

panfrost: Switch resources from an array to a set

This will help us reduce shared state and simplify multithreading, at
the expense of additional CPU overhead.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12528>

zink: stop referencing framebuffers

this is a waste of cycles now that surfaces are accurately tracked;
no-attachment fbs are still deferred to avoid premature deletion

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12429>

zink: defer deletion of no-attachment framebuffers

the ref on these is owned by the context, so defer deletion to avoid
premature destruction if the fb might be in use

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12429>

panfrost: Inline add_fbo_bos

Only used once, it's just complicating the batch cache interface.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12525>

panfrost: Remove get_fresh_batch

Unused, and of dubious value.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12525>

panfrost: Move bo->label assignment into the lock

We already took the lock, we just unlocked too early. Since the label is
reset in the BO cache, this is racy. Minimal impact in practice but is
still /wrong/ and caught by helgrind.

Fixes: 3fa1f93dace ("panfrost: Label all BOs in userspace")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12525>

panfrost: Don't use ralloc for resources

ralloc is not thread safe, so we cannot use a pipe_screen as a ralloc
context unless we lock the screen. The allocation patterns for resources
are trivial, so just use malloc/calloc/free directly instead of ralloc.
This fixes a segfault in:

dEQP-EGL.functional.sharing.gles2.multithread.random.images.copytexsubimage2d.1

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12525>

panfrost: Simplify get_fresh_batch_for_fbo

Makes the code easier to read, too.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12525>

panfrost: Remove null check in batch_cleanup

Shouldn't happen.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12525>

panfrost: Protect the variants array with a lock

Without a lock, two threads may bind the same shader CSO simultaneously,
allocate the same variant simultaneously, and then race each other in
the compiler. This manifests in various ways, most commonly failing the
assertion that UBO pushing has only run once. The simple_mtx_t solution
is used in Iris. Fixes the crash in:

dEQP-EGL.functional.sharing.gles2.multithread.simple.buffers.bufferdata_render

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12525>

panfrost/ci: Don't skip matrix inverse tests

Older versions of these tests were buggy and failed on Bifrost. The test
bug has been resolved upstream, but the skip list was not updated when
dEQP was uprevved with the fix. Run the tests.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12313>

panfrost/ci: Switch to suite support

Use the new deqp-runner suite support to combine our dEQP-GLES2,
dEQP-GLES3, and dEQP-GLES31 jobs into a single job. This simplifies load
balancing, enabling us to expand our test coverage without impacting
wall clock time.

With the new infrastructure in place, we add KHR-GLES* jobs for Mali
G52. This would have caught some recent regressions. Once we hit
conformance it's essential we remain conformant.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12313>

nir/gcm: pin some instructions which require uniform sources

fossil-db (Sienna Cichlid, GCM enabled):
Totals from 6192 (4.12% of 150170) affected shaders:
VGPRs: 548392 -> 542040 (-1.16%)
SpillSGPRs: 3702 -> 3990 (+7.78%); split: -0.54%, +8.32%
CodeSize: 62418488 -> 62481516 (+0.10%); split: -0.07%, +0.17%
MaxWaves: 70582 -> 71718 (+1.61%)
Instrs: 11768497 -> 11795079 (+0.23%); split: -0.07%, +0.30%
Latency: 445891848 -> 523561297 (+17.42%); split: -0.07%, +17.49%
InvThroughput: 115675481 -> 121494913 (+5.03%); split: -0.09%, +5.12%
VClause: 164914 -> 164934 (+0.01%); split: -0.05%, +0.06%
SClause: 405991 -> 395302 (-2.63%); split: -2.64%, +0.00%
Copies: 907216 -> 926429 (+2.12%); split: -1.11%, +3.23%
Branches: 456373 -> 457478 (+0.24%); split: -0.13%, +0.38%
PreSGPRs: 648030 -> 642953 (-0.78%); split: -0.88%, +0.10%
PreVGPRs: 522425 -> 516355 (-1.16%); split: -1.16%, +0.00%

Seems to affect Detroit: Become Human and Cyberpunk 2077. The Cyberpunk
2077 changes look like a fixed bug. At least some of the Detroit: Become
Human changes could probably be removed with better divergence analysis.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12444>

nir: consider push constant loads as always dynamically uniform

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timothy Arceri <tarceri@itsqueeze.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12444>

radv: call nir_lower_flrp() after the first radv_optimize_nir()

instead of inside the optimization loop

Totals from 2504 (1.67% of 150170) affected shaders: (GFX10.3)
VGPRs: 162592 -> 162416 (-0.11%); split: -0.12%, +0.01%
CodeSize: 18399756 -> 18383552 (-0.09%); split: -0.10%, +0.01%
MaxWaves: 42654 -> 42748 (+0.22%)
Instrs: 3499404 -> 3497075 (-0.07%); split: -0.08%, +0.01%
Latency: 87087238 -> 87064270 (-0.03%); split: -0.06%, +0.03%
InvThroughput: 21159621 -> 21150546 (-0.04%); split: -0.05%, +0.01%
VClause: 56653 -> 56667 (+0.02%); split: -0.00%, +0.03%
Copies: 226332 -> 226423 (+0.04%); split: -0.15%, +0.19%
Branches: 110027 -> 110025 (-0.00%); split: -0.05%, +0.04%
PreSGPRs: 168087 -> 168076 (-0.01%); split: -0.01%, +0.00%
PreVGPRs: 160814 -> 160705 (-0.07%)

Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12061>

nir/opt_algebraic: optimize flrp(fadd, fadd, x) only if fadd are used_once

Totals from 201 (0.13% of 150170) affected shaders: (GFX10.3)
VGPRs: 13880 -> 13856 (-0.17%)
CodeSize: 1517328 -> 1518124 (+0.05%); split: -0.04%, +0.10%
MaxWaves: 3184 -> 3192 (+0.25%)
Instrs: 285487 -> 285569 (+0.03%); split: -0.06%, +0.08%
Latency: 7774066 -> 7780877 (+0.09%); split: -0.10%, +0.19%
InvThroughput: 1936341 -> 1935287 (-0.05%); split: -0.07%, +0.02%
SClause: 11446 -> 11448 (+0.02%); split: -0.01%, +0.03%
Copies: 17500 -> 17506 (+0.03%); split: -0.51%, +0.55%
Branches: 8174 -> 8180 (+0.07%); split: -0.13%, +0.21%
PreVGPRs: 12507 -> 12427 (-0.64%)

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12061>

nir/loop_analyze: consider instruction cost of nir_op_flrp

Reviewed-by: Ian Romanick <ian.d.romanick@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12061>

venus: use uint32_t in vn_ring_submit

And in vn_ring_write_buffer as well, to fix the assert in
vn_ring_write_buffer.

The ring code uses 32-bit unsigned integers and relies on that their
overflow/underflow behavior is well-defined.  When ring->shared.head is
about to overflow and ring->cur has overflowed, this expression

  ring->cur + size - vn_ring_load_head(ring)

gives an incorrect result when size is 64-bit.

Signed-off-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Ryan Neph <ryanneph@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12494>

glthread: implement glGetUniformLocation without syncing

We already have the infrastructure for querying shader program properties
without syncing. This just uses it. _mesa_error_glthread_safe sets a GL
error from the producer thread.

This decreases CPU overhead for viewperf/snx.

Reviewed-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12490>

gallium: change pipe_draw_info::mode to uint8_t on MSVC to make it 1 byte large

needed by u_threaded_context

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12480>

gallium: use a packed enum to make pipe_prim_mode 1-byte large with __GNUC__

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12480>

zink: only force all buffer rebinds if rebinds exist on other contexts

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12425>

zink: rebind all buffers on replacement

in theory it's possible to trigger cases where rebinds aren't based on the
current context, so ensure that (very unlikely) case is handled

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12425>

zink: count streamout rebinds when doing buffer rebinds

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12425>

zink: add bind counts for so bindings

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12425>

zink: split out buffer rebinds to helper functions

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12425>

zink: make descriptor update functions return the updated resource

convenience++

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12425>

zink: clear out all ubo rebinds first if they exist

these are the second most common rebind, so iterate them all first

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12425>

radv: use nir_vector_insert_imm in lower_intrinsics

This creates a single nir_op_vecn instead of a nir_op_vecn and several
copies.

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/12469>

nir/lower_io: use nir_vector_insert_imm()

This creates a single nir_op_vecn instead of a nir_op_vecn and several
copies.

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/12469>

gitlab-ci: Fix octopus device type and tag

* This month, the octopus device was renamed on Collabora's LAVA lab, so
  we need to update it to get it working again on Mesa CI.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12507>

Revert "nir/opt_algebraic: optimize fmax(-fmin(b, a), b) -> fmax(b, -a)"

This is wrong for negative values.

This reverts commit 07cd30ca293d1eb6980f69f330f9d182652cf902.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12515>

lavapipe: fix reported subpixel precision for lines

We have no reason to report a subpixel precision of 4 for lines; in fact
LLVMpipe uses 8 subpixel bits for lines, similar to other primitives.
But let's use the pipe-cap for this instead of hard-coding it.

Fixes: 9fbf6b2abff ("lavapipe: implement VK_EXT_line_rasterization")
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12277>

broadcom/compiler: Fix qpu.flags.muf typo.

Fix defect reported by Coverity Scan.

Same on both sides (CONSTANT_EXPRESSION_RESULT)
pointless_expression: The expression inst->qpu.flags.auf !=
V3D_QPU_UF_NONE || inst->qpu.flags.auf != V3D_QPU_UF_NONE does not
accomplish anything because it evaluates to either of its identical
operands, inst->qpu.flags.auf != V3D_QPU_UF_NONE.

Fixes: 3f2c54a27ff ("broadcom/compiler: rewrite partial update liveness tracking")
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Reviewed-by: Juan A. Suarez <jasuarez@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12385>

llvmpipe: improve polygon-offset precision

This performs the polygon offset addition after interpolation, which
prevents floating-point cancellation issues completely.

This does mean that we have to perform a single floating-point addition
more per fragment than before, unless we also want to spend a bit in
the fragment-shader variant key to avoid this.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12442>

llvmpipe: split coefficient calculation and store

This will be used for some underhanded smuggling of values in the next
commit.

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12442>

llvmpipe: clamp z to 0..1 range when using polygon offset

The OpenGL 4.6 compatibility spec, section 14.6.5 (Depth Offset) says
the following:

> For fixed-point depth buffers, fragment depth values are always
> limited to the range [0,1] by clamping after offset addition is
> performed. Fragment depth values are clamped even when the depth
> buffer uses a floating-point representation.

So we need to properly clamp the result here.

This fixes the following dEQP failures, that the CI has missed:

- dEQP-GLES3.functional.polygon_offset.default_result_depth_clamp
- dEQP-GLES3.functional.polygon_offset.default_factor_1_slope
- dEQP-GLES3.functional.polygon_offset.fixed24_result_depth_clamp
- dEQP-GLES3.functional.polygon_offset.fixed24_factor_1_slope

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12442>

crocus: copy views before adjusting

The current code overwrote the original view which meant if we
had to reemit a surface the second emit would be wrong.

This fixes cubemaps on gm45 and maybe some issues with 3D textures
elsewhere.

Fixes: f3630548f1da ("crocus: initial gallium driver for Intel gfx 4-7")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12514>

freedreno: Require C++17.

Commit 3a772be026c ("freedreno: Add perfetto renderpass support")
uses C++17 init-statement feature.

GCC
../src/gallium/drivers/freedreno/freedreno_perfetto.cc: In lambda function:
../src/gallium/drivers/freedreno/freedreno_perfetto.cc:148:11: warning: init-statement in selection statements only available with ‘-std=c++17’ or ‘-std=gnu++17’
  148 |       if (auto state = tctx.GetIncrementalState(); state->was_cleared) {
      |           ^~~~

Clang
../src/gallium/drivers/freedreno/freedreno_perfetto.cc:148:11: warning: 'if' initialization statements are a C++17 extension [-Wc++17-extensions]
      if (auto state = tctx.GetIncrementalState(); state->was_cleared) {
          ^

Intel C++ Compiler
../src/gallium/drivers/freedreno/freedreno_perfetto.cc(148): error: expected a ")"
        if (auto state = tctx.GetIncrementalState(); state->was_cleared) {
                                                   ^

Fixes: 3a772be026c ("freedreno: Add perfetto renderpass support")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/5193
Suggested-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Signed-off-by: Vinson Lee <vlee@freedesktop.org>
Acked-by: Rob Clark <robdclark@chromium.org>
Acked-by: Erik Faye-Lund <erik.faye-lund@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12293>

intel/compiler: Add unified barrier support for CS

Program CS barrier message fields for producers/consumers.

Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11963>

intel/compiler: Add unified barrier support for TCS

Program the producers/consumer fields for TCS Barrier messages.
Producer and consumer fields are set to number of TCS threads.

Ref: Bspec 54006 for Barrier Data Payload
Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11963>

intel/compiler: Regroup TCS barrier code paths

Rearrange if/else fragments to unify case for Gen11 or later
platforms. This will help the code look cleaner for adding
unified barrier support to TCS.

Signed-off-by: Jordan Justen <jordan.l.justen@intel.com>
Reviewed-by: Sagar Ghuge <sagar.ghuge@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11963>

panfrost: Rip out primconvert code

This is handled in common Gallium code if we set the appropriate CAP.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Suggested-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12509>

panfrost: Fix NULL dereference in allowlist code

If a user attempts to run Panfrost on an unsupported GPU (e.g. Mali
T604), Panfrost will refuse to load and will destroy the screen
immediately, allowing for a graceful fallback to a software rasterizer.
However, the screen destroy code calls a screen_destroy function in the
GenXML vtbl -- and this function is still NULL when the allowlist is
checked. This manifests as crashes on unsuported GPUs.

Issue tracked down with Icecream95's mad Ghidra skills.

Closes: #5269
Fixes: 88dc4db6be7 ("panfrost: Init/destroy blitter from per-gen file")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Reported-by: Icecream95 <ixn@disroot.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12512>

intel: Parse INTEL_NO_HW for devinfo construction

This commit does several things:

* Unify code common to several drivers by evaluating INTEL_NO_HW within
  intel_get_device_info_from_fd (suggested by Jordan).
* For drivers that keep a copy of the intel_device_info struct, a
  separate copy of the no_hw field is now unnecessary. Remove them.
* Minimize kernel queries when INTEL_NO_HW is true. This is done for
  code simplification, but we may find reason to undo this later on.

Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12007>

intel: Use env_var_as_boolean for INTEL_NO_HW

The prior method of checking the result of getenv() for NULL would cause
the feature to be enabled for INTEL_NO_HW=0.

Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Caio Marcelo de Oliveira Filho <caio.oliveira@intel.com>
Reviewed-by: Jordan Justen <jordan.l.justen@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12007>

panfrost: Port v5 blend shader issue to blitter

This is a presumed erratum workaround. Fixes INSTR_INVALID_PC faults on
some draw_buffers_indexed.* cases on Midgard, where a blend shader is
required to pack RT n > 0.

Backport the workaround from the GL driver. The helper is now in common
code for panvk to use as well; it has the same bug.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

panfrost: Zero initialize blend_shaders

Fixes an invalid read caught by valgrind when there is a hole in the
valid render target mask:

==6749== Conditional jump or move depends on uninitialised value(s)
==6749==    at 0x5E88EC0: panfrost_prepare_fs_state (pan_cmdstream.c:417)
==6749==    by 0x5E88EC0: panfrost_emit_frag_shader (pan_cmdstream.c:501)
==6749==    by 0x5E88EC0: panfrost_emit_frag_shader_meta (pan_cmdstream.c:573)
==6749==    by 0x5E88EC0: panfrost_update_state_fs (pan_cmdstream.c:2593)
==6749==    by 0x5E8B0BF: panfrost_direct_draw (pan_cmdstream.c:2839)

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Fixes: a124c47b9f9 ("panfrost: Fix NULL derefs in pan_cmdstream.c")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/mdg: Handle swapped 565 and 1010102 unorm

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/lower_framebuffer: Don't open-code pan_unpacked_type_for_format

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/lower_framebuffer: Don't open-code pad_vec4

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/lower_framebuffer: Don't treat UNORM 4 special

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/lower_framebuffer: Unify UNORM handling

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/lower_framebuffer: Use fmul_imm

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/lower_framebuffer: Don't replicate so much

We need to replicate to deal with multisampling, but not otherwise.
Simplify the logic substantially.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

pan/mdg: Insert moves before writeout when needed

Otherwise we end up accessing overwritten registers. Fixes

dEQP-GLES31.functional.draw_buffers_indexed.overwrite_common.common_enable_buffer_enable

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

panfrost: Delete unpacks for blendable formats

Unnecessary.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

panfrost: Use blendable check for tib read check

These are the same! Either you're blendable and can use f32/f16
conversion, or you're raw and you can only get raw. It's that simple!

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

panfrost: Fix UNORM 10 sizes

Fixes: 56047fb64d7 ("panfrost: Fix UNORM 16 rendering")
Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

panfrost: Remove unneeded quirks from T760

Will cause trouble later in the series when we start garbage collecting
unneeded code.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

panfrost: Add explicit padding to pan_blend_shader_key

So the hash function doesn't end up hashing uninitialized values.

Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Reported-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Fixes: bbff09b9521f ("panfrost: Move the blend shader cache at the device level")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

panfrost: Add padding to pan_blit_blend_shader_key

So the hashtable helpers know the correct size of the struct.

Signed-off-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11383>

iris: Mark the aux table buffers with EXEC_OBJECT_CAPTURE.

Having these could be useful when tracking down GPU hangs.

Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12420>

iris: Bypass the BO cache when allocating buffers for aux map tables

When freeing a buffer, we may return a non-idle buffer to the cache,
which means we cannot unmap aux entries at that time.  Instead, we
defer unmapping the stale aux entry until we reuse a BO from the cache.

Unfortunately, this can lead to a recursive locking issue:

1. intel_aux_map_add_mapping wants to set up a new aux entry

   It takes the intel_aux_map_context::mutex lock, then calls:

   add_mapping -> get_aux_entry -> add_sub_table -> add_buffer ->
   intel_aux_map_buffer_alloc -> iris_bo_alloc

2. iris_bo_alloc tries to allocate a BO from the cache, doing:

   alloc_bo_from_cache -> intel_aux_map_unmap_range ->
   intel_aux_unmap_range

   ...which then tries to take the intel_aux_map_context::mutex lock.
   But it is already locked.

One solution would be to rework the aux map handling code to allocate
BOs without holding its lock, but that looks to be painful.  Another
is to make the lock recursive, but we try and avoid that.  A third
option wuold be to add a BO_ALLOC flag that makes alloc_bo_from_cache
skip any buffers with aux_map_address != 0 so we don't have to unmap,
making the less cache effective but fixing the recursive lock.

A fourth option is to simply bypass the BO cache altogether for the
buffers that hold the aux map itself.  Allocating new BOs for the aux
tables should be relatively rare, so there's probably not a lot of
benefit in using the BO cache.

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/5191
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12420>

venus: scrub ignored fields of pipeline info when rasterization is disable

v2: use vk_alloc instead of vk_zalloc because of full memcpy

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com> (v1)
Reviewed-by: Ryan Neph <ryanneph@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12499>

venus: fix all missing vn_object_base_fini

Signed-off-by: Yiwei Zhang <zzyiwei@chromium.org>
Reviewed-by: Chia-I Wu <olvaffe@gmail.com>
Reviewed-by: Ryan Neph <ryanneph@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12498>

tu: Enable VK_KHR_uniform_buffer_standard_layout

This extension relaxes the alignment requirements to allow the GL std430
layout to be used. freedreno/ir3 already supports this (via
PIPE_CAP_LOAD_CONSTBUF).

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12495>

nir/opt_algebraic: optimize fmax(-fmin(b, a), b) -> fmax(b, -a)

Found with Cyberpunk 2077.

fossils-db (GFX10.3):
Totals from 128 (2.34% of 5465) affected shaders:
CodeSize: 769720 -> 767656 (-0.27%); split: -0.27%, +0.00%
Instrs: 145748 -> 145229 (-0.36%)

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Daniel Schürmann <daniel@schuermann.dev>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11604>

vulkan/wsi/sw: wait for image fence before submitting to queue

With hw devices, when you submit a present, implicit sync will
make sure the work submitted to the gpu on the client will end
up happening before the present work submitted on the server.

However with sw paths there is no real GPU, the lavapipe fake
GPU thread is client side only and presenting is done directly
from the pixmap (or later shared pixmap). In order for this to
make sense the wsi common code should wait for the fence on the
image before queueing the submit to the server so that all
client works has been flushed to the pixmap before the copy or
present operation is submitted.

Fixes: 8004fa9c9501 ("vulkan/wsi: add sw support. (v2)")
Acked-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12502>

aco/scheduler: allow moving down VMEM stores to below VMEM loads

fossil-db (Vega10):
Totals from 93 (0.06% of 150305) affected shaders:
SGPRs: 4832 -> 4768 (-1.32%)
VGPRs: 4084 -> 4144 (+1.47%)
CodeSize: 316080 -> 317208 (+0.36%); split: -0.11%, +0.47%
MaxWaves: 589 -> 580 (-1.53%)
Instrs: 60229 -> 60511 (+0.47%); split: -0.15%, +0.61%
Latency: 636477 -> 540029 (-15.15%); split: -15.26%, +0.10%
InvThroughput: 293027 -> 283043 (-3.41%); split: -4.21%, +0.80%
VClause: 2557 -> 2716 (+6.22%); split: -0.35%, +6.57%
SClause: 1381 -> 1395 (+1.01%); split: -0.14%, +1.16%
Copies: 9424 -> 9728 (+3.23%); split: -0.74%, +3.97%

fossil-db (Sienna Cichlid):
Totals from 88 (0.06% of 150170) affected shaders:
VGPRs: 3840 -> 3872 (+0.83%)
CodeSize: 300544 -> 300960 (+0.14%); split: -0.09%, +0.23%
Instrs: 53714 -> 53871 (+0.29%); split: -0.05%, +0.35%
Latency: 489854 -> 462001 (-5.69%); split: -6.30%, +0.61%
InvThroughput: 100307 -> 95142 (-5.15%); split: -5.50%, +0.35%
VClause: 2322 -> 2564 (+10.42%); split: -0.39%, +10.81%
SClause: 1345 -> 1358 (+0.97%); split: -0.15%, +1.12%
Copies: 4113 -> 4351 (+5.79%); split: -0.66%, +6.44%

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/12211>

llvmpipe: use preferred attribute interpolation for wide lines

When rasterizing legacy-lines, OpenGL defines the width as being an
extrusion along the minor axis, repeating varyings. While the spec
*does* allow for an alternative method that matches our current results,
the OpenGL ES CTS doesn't allow these results even if OpenGL ES has the
same wording of an alternative method.

This is technically speaking a bug in the OpenGL ES CTS, but it seems
like nobody else is using the alternative formulation, at least not
while passing the OpenGL ES CTS. On top of this, the OpenGL specification
explicitly lists the extrusion results as the preferred method.

So it seems like a good idea for us to do this the way the OpenGL
specification prefers regardless; it's going to give less surprising
results to applications, and it's helping us pass some tests.

This math to set these up would "trivially" be:

dx = (dx * dx + dy * dy) / dx
dy = 0

and:

dy = (dx * dx + dy * dy) / dy
dx = 0

...but since we've already calculated dxdy, we can reformulate this to
save a division.

This fixes the following dEQP test-cases:
- dEQP-GLES2.functional.rasterization.interpolation.basic.line_loop_wide
- dEQP-GLES2.functional.rasterization.interpolation.basic.line_strip_wide
- dEQP-GLES2.functional.rasterization.interpolation.basic.lines_wide
- dEQP-GLES2.functional.rasterization.interpolation.projected.line_loop_wide
- dEQP-GLES2.functional.rasterization.interpolation.projected.line_strip_wide
- dEQP-GLES2.functional.rasterization.interpolation.projected.lines_wide
- dEQP-GLES3.functional.rasterization.fbo.rbo_singlesample.interpolation.lines_wide
- dEQP-GLES3.functional.rasterization.fbo.texture_2d.interpolation.lines_wide
- dEQP-GLES3.functional.rasterization.interpolation.basic.line_loop_wide
- dEQP-GLES3.functional.rasterization.interpolation.basic.line_strip_wide
- dEQP-GLES3.functional.rasterization.interpolation.basic.lines_wide
- dEQP-GLES3.functional.rasterization.interpolation.projected.line_loop_wide
- dEQP-GLES3.functional.rasterization.interpolation.projected.line_strip_wide
- dEQP-GLES3.functional.rasterization.interpolation.projected.lines_wide

Reviewed-by: Roland Scheidegger <sroland@vmware.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11315>

aco: remove label_extract if the extract is used by a non-VALU

If an extract is used by a non-VALU instruction, it can't be applied to
all instructions, so it's not beneficial to try to apply it.

This check isn't needed because can_apply_extract()/can_use_SDWA() should
already handle non-VALU instructions.

fossil-db (Sienna Cichlid):
Totals from 1020 (0.68% of 150170) affected shaders:
SpillSGPRs: 1577 -> 1571 (-0.38%)
CodeSize: 7863668 -> 7858336 (-0.07%); split: -0.07%, +0.00%
Instrs: 1431583 -> 1431083 (-0.03%); split: -0.04%, +0.01%
Latency: 25891250 -> 25890916 (-0.00%); split: -0.01%, +0.01%
InvThroughput: 7248683 -> 7248655 (-0.00%); split: -0.01%, +0.01%
SClause: 49072 -> 49071 (-0.00%)
Copies: 126649 -> 126580 (-0.05%); split: -0.11%, +0.06%
Branches: 39129 -> 39120 (-0.02%); split: -0.03%, +0.01%
PreSGPRs: 53071 -> 52943 (-0.24%); split: -0.26%, +0.02%
PreVGPRs: 57437 -> 57435 (-0.00%); split: -0.01%, +0.01%

fossil-db (Polaris10):
Totals from 654 (0.43% of 151696) affected shaders:
CodeSize: 5814552 -> 5811568 (-0.05%); split: -0.05%, +0.00%
Instrs: 1105783 -> 1105049 (-0.07%); split: -0.07%, +0.00%
Latency: 20261458 -> 20259744 (-0.01%); split: -0.01%, +0.00%
InvThroughput: 9011785 -> 9011749 (-0.00%); split: -0.00%, +0.00%
Copies: 104693 -> 103904 (-0.75%); split: -0.76%, +0.00%
PreSGPRs: 36105 -> 36095 (-0.03%); split: -0.03%, +0.01%
PreVGPRs: 43813 -> 43809 (-0.01%)

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/12212>

radv: allocate shaders to 32-bit address to skip PGM_HI

This reduces the number of emitted registers.

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/12466>

radv: don't use SQ_NON_EVENT before GE_PC_ALLOC for better perf on Navi1x

Seems it make the perf worse.
Ported from RadeonSI.

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/12466>

aco: add more validation rules for SDWA operands

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco/opcodes: remove definition_size[]

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco/validate: simplify get_subdword_bytes_written()

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco/ra: refactor subdword operand stride

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco/ra: refactor subdword definition info

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco: add instr_is_16bit() helper function

to indicate whether some instruction writes partial registers, only.

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco: use VOPC_SDWA on GFX9+

Totals from 5138 (3.42% of 150170) affected shaders: (GFX10.3)
VGPRs: 409520 -> 409416 (-0.03%); split: -0.03%, +0.00%
CodeSize: 43056360 -> 43035696 (-0.05%); split: -0.06%, +0.02%
MaxWaves: 69296 -> 69310 (+0.02%)
Instrs: 8161016 -> 8153365 (-0.09%); split: -0.10%, +0.01%
Latency: 109397002 -> 109756208 (+0.33%); split: -0.05%, +0.38%
InvThroughput: 23238920 -> 23310761 (+0.31%); split: -0.11%, +0.42%
VClause: 135141 -> 135100 (-0.03%); split: -0.05%, +0.02%
SClause: 349511 -> 349489 (-0.01%); split: -0.01%, +0.00%
Copies: 388107 -> 387754 (-0.09%); split: -0.48%, +0.38%
Branches: 184629 -> 184503 (-0.07%); split: -0.08%, +0.01%
PreSGPRs: 258807 -> 258839 (+0.01%)
PreVGPRs: 372561 -> 372184 (-0.10%); split: -0.10%, +0.00%

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco/print_ir: fix printing of VOPC_SDWA definitions

Reviewed-by: Rhys Perry <pendingchaos02@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12364>

aco: fix vectorized 16-bit load_input/load_interpolated_input

Seems we haven't encountered this before because
nir_lower_io_to_scalar_early usually scalarizes this.

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/12486>

radv: remove useless DISABLE_{ZMASK,SMEM}_EXPCLEAR_OPTIMIZATION state

This has no effect without enabling EXPCLEAR.

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/12326>

radv: remove unused fast depth-stencil gfx clear path with expclear

This has never been used because it requires to know the previous
clear values which is not really possible in Vulkan.

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/12326>

lavapipe: fix missing VKAPI_CALL attribute

Fixes build on mingw

Fixes: c198adf7
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12484>

util/xmlconfig: Test values set via the environment

driconf options can also be set via environment variables.  This is a
simple touch-test of that feature.

Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12477>

util/xmlconfig: Make unit tests more resilient against user env settings

Before this, setting 'vblank_mode=0' in the environment would cause a
unit test to fail.

Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12477>

frontend/dri: add environment variable DRI_NO_MSAA for performance comparisons

Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12491>

radeonsi: remove vertices_per_patch parameter from draw-related functions

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12351>

gallium: remove vertices_per_patch, add pipe_context::set_patch_vertices

We would like draw-only display lists to have immutable draw info and
this is the only GL non-draw state in pipe_draw_info (not counting
view_mask).

It also allows removing some code from draw_vbo for tessellation.

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12351>

tu: Remove some stale bypass xfails

These were fixed by 09e0b29bb63f60231b26b4c8f02eadb68e51b623 which was
missed during the suite conversion. For the remaining still-valid fail,
there is a CTS patch in progress.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12488>

freedreno/crashdec: Quiet spammy print in query mode

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12489>

freedreno/crashdec: Decode full RB in verbose mode

This is useful to get a better view of previous commands in the
ringbuffer.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12489>

freedreno/cffdec: Fix gpuaddr comparision

gpuaddrs are 64b, and they can be more than 2^^32 apart.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12489>

freedreno/cffdec: Fix indentation

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12489>

pan/bi: Extend bi_add_nop_for_atest for tilebuffer loads

Fixes framebuffer_fetch and blend_equation_advanced dEQP tests on v6.

v2: Use clause dependencies rather than comparing the message type
v3: Shift the BIFROST_SLOT_* constants before using them as a mask

Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12375>

tu: Free device->bo_idx and device->bo_list on init failure

Two related changes:

- in tu_device.c:tu_CreateDevice we need to free both pointers in the
  teardown path after tu_bo_finish(global_bo), which uses the pointers.
  They are allocated in the first call to tu_bo_init(), which happens
  when global_bo is allocated.

- in tu_drm.c:tu_bo_init we need to free bo_list if the bo_idx
  allocation fails. Convert to the goto teardown pattern as well.

Fixes the following dEQP-VK tests:
  dEQP-VK.api.device_init.create_instance_device_intentional_alloc_fail
  dEQP-VK.api.object_management.alloc_callback_fail.device
  dEQP-VK.api.object_management.alloc_callback_fail.device_group

Reviewed-by: Kristian H. Kristensen <hoegsberg@google.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12481>

pan/bi: Use CLPER_V6 on Mali G31

Apparently, CLPER_V7 is missing from Mali G31, but CLPER_V6 works. Fixes
INSTR_INVALID_ENC faults and failures in
dEQP-GLES3.functional.shaders.derivate.* on Dvalin.

Technically not an errata but an implementation difference. I suspect
Mali G51 will need this as well, should we ever allowlist it.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Cc: mesa-stable
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12478>