ci/windows: Windows runner is back online

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17962>

anv: emit scissors when the pipeline changes

With the switch to common dynamic state tracking, something got lost
that made the scissors not always be emitted when they are not dynamic
and the pipeline is marked dirty.

Since both viewport and scissors make use of each other to calculate
their values, just stick the scissor emit in the same if block as
viewport for now.
I'd rather have them decoupled, and at least the Vulkan CTS didn't
complain when I tried it, but I don't know what other effects that
may have, especially when it comes to the guardband.

Fixes a bunch of tests under
dEQP-VK.pipeline.*.multisample.misc.*

Fixes: 7d25c04236b ("anv: Switch to using common dynamic state tracking")

Reviewed-by: Jason Ekstrand <jason.ekstrand@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17964>

ci/lavapipe: wine flakes, switch to manual for now

Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7030

Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17965>

gallium/u_blitter: make the bilinear filter for MSAA resolving conformant

This implements the bilinear filter exactly like OpenGL.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

gallium/u_blitter: make nearest filtering emulation using TXF conformant

This is required by glBlitFramebuffer.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Acked-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

gallium/u_blitter: clean up IMMs in util_make_fs_blit_msaa_gen

Always declare IMM[0].

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

gallium/u_blitter: treat Z scaling as scaled blits

Hopefully this didn't break anything.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

gallium/u_blitter: fix the has_txf support condition

GLSL 1.30 has it too.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

gallium/u_blitter: remove unused code for integer MSAA resolve

GL doesn't have it.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

gallivm: initialize texture_unit_offset in emit_size_query

this fixes a crash, suggested by Dave Airlie

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

gallivm: fix TXQ crash with MSAA samplers

suggested by Dave Airlie

Reviewed-by: Dave Airlie <airlied@redhat.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

ttn: set the correct sampler declaration type in the presense of txs and lod

We used the result type of lod and txs for sampler declarations,
which broke following instructions that are not lod and txs.

Use the sampler type from TGSI if it's present, not the result type
of lod and txs.

Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17692>

ci/tu+zink: Re-enable the traces that were asserting about UBWC compatibility.

The tu and zink fixes have made them work now.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17900>

tu: Use the format list to decide whether we can do UBWC with MUTABLE.

The goal here is to allow zink and ANGLE to switch sRGB on and off, but
keep UBWC enabled for most textures.

This doesn't quite work yet for zink on gbm, where zink asks the device if
it supports a UBWC-modifier image with unorm+sRGB mutable formats (yes),
but then doesn't pass the unorm+sRGB format list to vkImageCreate(), which
then asserts because it couldn't make a general
VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT UBWC image.

The format list here comes from Danylo's testing in
https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17278#note_1447812

Reduces frame times in Genshin Impact GLES AoV 2.02063% +/- 0.207575% (n=5)

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17900>

tu: Treat viewFormatCount==0 as no format list.

"If viewFormatCount is zero, pViewFormats is ignored and the image is
 created as if the VkImageFormatListCreateInfo structure were not included
 in the pNext chain of VkImageCreateInfo."

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17900>

tu: Move the vkCreateImage format list checks to helper functions.

Mixing up the two checks made them both harder to read.

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17900>

zink: Make sure that we keep the existing ici pNext chain on inserts.

For external image imports, we'd lose the mutable image format list,
causing turnip to get angry that we try to do UBWC despite not having a
UBWC-compatible format list.

Cc: mesa-stable
Fixes: 28ee911ad671 ("zink: handle mutable swapchain images with dmabuf")
Reviewed-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17900>

c11: Fixes the declaration of timespec_get in time.h for macOS

../src/util/tests/timespec_test.cpp:327:4: error: use of undeclared identifier 'timespec_get'; did you mean 'timespec_eq'?
   timespec_get(&a, TIME_UTC);
   ^~~~~~~~~~~~
   timespec_eq
../src/util/timespec.h:299:1: note: 'timespec_eq' declared here
timespec_eq(const struct timespec *a, const struct timespec *b)

Signed-off-by: Yonggang Luo <luoyonggang@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16995>

util: open_memstream also support on MacOS

Signed-off-by: Yonggang Luo <luoyonggang@gmail.com>
Acked-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/16995>

microsoft/clc: Fixes compiling errors with clang/mingw64 in clc/clc_compiler_test.cpp

clc_compiler_test.cpp:1322:67: error: non-constant-expression cannot be narrowed from type 'double' to 'float' in initializer list
      log(0.0f) / log(2), log(1.0f) / log(2), log(2.0f) / log(2), log(3.0f) / log(2)
clc_compiler_test.cpp:2306:25: error: non-constant-expression cannot be narrowed from type 'std::vector<unsigned int>::size_type' (aka 'unsigned long long') to 'unsigned int' in initializer list
   CompileArgs args = { inout.size(), 1, 1 };

Signed-off-by: Yonggang Luo <luoyonggang@gmail.com>
Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17926>

meson: clc compiler test only works on win32

Signed-off-by: Yonggang Luo <luoyonggang@gmail.com>
Acked-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17934>

radv: fix gathering XFB info if there is dead outputs

The driver should still gather XFB info even if all XFB outputs are
dead, otherwise the pipeline can't find the streamout shader.

RADV should use vk_spirv_to_nir() at some point to reduce code
duplication during SPIRV->NIR compilation.

This fixes new dEQP-VK.transform_feedback.simple.*.

Cc: mesa-stable
Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-by: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17939>

radv/ci: Put one more board to run the CTS on Stoney Ridge

These jobs were taking more than 20 minutes to run, slowing the pipeline
down.

Reviewed-by: David Heidelberg <david.heidelberg@collabora.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17935>

radeonsi/ci: Move libva and piglit jobs to Raven Ridge boards

As we have some free, and then we can reassign two Stoney Ridge boards
to reduce the duration for some jobs that are taking too long currently.

Also, these machines are much faster so we are able to run all of the
Piglit tests and a bigger portion of the VA tests.

Reviewed-by: David Heidelberg <david.heidelberg@collabora.com>
Reviewed-by: Emma Anholt <emma@anholt.net>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17935>

radeonsi: print radeon_bo_flag when AMD_DEBUG=vm is used

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11449>

radeonsi: allocate BIND_RENDER_TARGET as tmz instead of SCANOUT

Using AMD_DEBUG=tmz to enable tmz without the application asking
for it is useful for testing. Since SCANOUT isn't set on Wayland,
use RENDER_TARGET instead as a heuristic to try to allocate color
framebuffer as tmz.

With this change we can enable tmz in x11 and wayland.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11449>

amdgpu/bo: update uses_secure_bos when importing buffers

Fixes: 90b98c06493 ("amd/tmz: move uses_secure_bos to radeon_winsys")
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11449>

gallium/dri2: make protected buffer checks opt-in

This check is an infinite source of issues when testing protected
content, so make it opt-in (eg for developers of applications dealing
with protected content buffers).

On AMD and Intel importing a protected buffer as non-protected will cause
the display to be incorrect (= garbage).

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/11449>

ci/windows: Disable Windows builds while the runner is offline again

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17958>

docs/ci: show how to run traces locally

Newcomers may want to test traces locally and understand how the replay
works.

Acked-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Signed-off-by: David Heidelberg <david.heidelberg@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17347>

radv/ci: bump the console activity timeout of VanGogh to 3 minutes

It seems like VKCTS's startup time is close to the current console
activity timeout (2 minutes...) on the Steam Deck, leading to spurious
timeouts...

Let's bump the timeout by 50%!

Signed-off-by: Martin Roukala (né Peres) <martin.roukala@mupuf.org>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17829>

radv/ci: skip host_write_transfer_src.1048576 on renoir

The test can take a little longer than 60s, which leads to spurious
timeouts and thus unstable CI pipelines.

Signed-off-by: Martin Roukala (né Peres) <martin.roukala@mupuf.org>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17829>

radv/ci: add more tests to the flake list of vega10

Signed-off-by: Martin Roukala (né Peres) <martin.roukala@mupuf.org>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17829>

radv/ci: move some tests from the vega10 fail to its flake list

v2:
 - Move the other 2 `sample_texture` tests to the flake list (Ishi)

Signed-off-by: Martin Roukala (né Peres) <martin.roukala@mupuf.org>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17829>

radv/ci: document a recent regression

This regression may have been introduced by mesa/mesa!17652 or
mesa/mesa!17625.

Signed-off-by: Martin Roukala (né Peres) <martin.roukala@mupuf.org>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17829>

freedreno: Cap reported video memory by VA size

Don't report that we have more video memory than the GPU's virtual
address space size.

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17947>

freedreno/drm: Add FD_VA_SIZE param

Signed-off-by: Rob Clark <robdclark@chromium.org>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17947>

dzn: Use D3D12 constants instead of naked numbers for device limits

Reviewed-by: Giancarlo Devich <gdevich@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17915>

dzn: Disable extensions that require multi-view

Reviewed-by: Giancarlo Devich <gdevich@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17915>

docs: fixup link to virgl docs

Fixes: 6897266ce01 ("docs: import virgl docs")
Acked-by: Chia-I Wu <olvaffe@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17881>

wgl: Release the context before deleting it in wglDeleteContext

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17848>

wgl: Always release local references in stw_make_current_by_handles

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17848>

wgl: Homogenize read framebuffer reference counting logic

Reviewed-by: Jesse Natalie <jenatali@microsoft.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17848>

radeonsi: use do..while loops and other cosmetic changes in display list path

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17933>

radeonsi: use si_cp_dma_prefetch_inline for prefetching VBO descriptors

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17933>

radeonsi: remove temporary si_context::vb_descriptor_user_sgprs

We were writing descriptors into si_context and then copying them into
the command buffer. Just write them into the command buffer directly.
Also set the pointer to VBO descriptors right after them.

When we start a new command buffer or we finish blitting, we no longer
restore precomputed VBO descriptors. Instead, we just reupload them again.
It's a compromise to have the common path simpler and faster (maybe).

This removes a lot of stuff. Now the VBO descriptor upload path looks
very similar to the display list path.

There was an accidental hidden optimization that is now documented as
"last_const_upload_buffer".

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17933>

radeonsi: remove vb_descriptors_gpu_list only used for debugging

While this is nice to have, it doesn't include VBO descriptors in user
SGPRs, and we need to remove it, so that we can simplify the VBO code.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17933>

radeonsi: add vertex buffers into the BO list in set_vertex_buffers

This is more straightforward. Also, radeon_add_to_buffer_list makes
writing VBO descriptors into the command buffer slower after that code
is reordered in following commits. This seems to be the only way that
isn't slower.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17933>

radeonsi: merge both fail paths in si_set_vb_descriptor

I removed the assertion because apps are allowed to set an offset greater
than the size.

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17933>

tu: Fix sysmem depth attachment clear flushing

We can't invalidate CCU if there is any dirty data that hasn't been
flushed yet. In the case where we clear depth, we know that the depth
attachment itself isn't dirty but there may be dirty data from other
renderpasses. Therefore we need to flush before invalidating depth.

Fixes: 487aa80 ("tu: Rewrite flushing to use barriers")
Closes: #6987
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17940>

docs: document DRI_PRIME

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17298>

vulkan/device_select: allow DRI_PRIME=vendor_id:device_id

To match the GL side.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17298>

vulkan/device_select: print the dri_prime warning only if needed

The next commit will allow a different DRI_PRIME syntax, so move
this printf in the right if block.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17298>

loader: allow DRI_PRIME=vendor_id:device_id syntax

This syntax allows to select a specific GPU without depending on
the pci bus information.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17298>

loader: don't return empty string in loader_get_dri_config_device_id

The caller expects a NULL return value if the option isn't set.

Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17298>

ci/windows: Re-enable Windows runners

Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17942>

radv: simplify radv_bind_dynamic_state() slightly

This adds RADV_CMP_COPY to compact copies. Based on ANV.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17649>

radv: remove unused states parameter from some radv_emit_XXX() helpers

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17649>

radv: simplify saving/restoring all dynamic states

Instead of copying every field individually, just use a whole memcpy.
This could be optimized but that's not the point here.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17649>

radv: remove RADV_META_SAVE_SAMPLE_LOCATIONS

We already save/restore all other dynamic states unconditionally, it's
not really useful to make an exception for sample locations.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17649>

aco: improve VcmpxPermlaneHazard workaround

According to LLVM, we only need to care about VOPC which writes exec.

No fossil-db changes.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17697>

aco: only add vscnt wait when visiting VMEM/DS

This prevents issues where we insert a s_waitcnt_vscnt(0) at the start of
a block or very end of the shader because we're joining two blocks (for
example, one with has_VMEM=true and the other with
has_branch_after_DS=true).

fossil-db (navi10):
Totals from 2441 (1.51% of 161220) affected shaders:
Instrs: 1383964 -> 1384094 (+0.01%); split: -0.07%, +0.08%
CodeSize: 7438212 -> 7438760 (+0.01%); split: -0.05%, +0.06%
Latency: 13780665 -> 13679664 (-0.73%); split: -1.53%, +0.80%
InvThroughput: 2950835 -> 2921511 (-0.99%); split: -1.06%, +0.07%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17697>

aco: set has_VMEM,has_DS=false after a branch

fossil-db (navi10):
Totals from 161 (0.10% of 161220) affected shaders:
Instrs: 206726 -> 207179 (+0.22%); split: -0.02%, +0.24%
CodeSize: 1114152 -> 1116032 (+0.17%); split: -0.01%, +0.18%
Latency: 2119380 -> 2147403 (+1.32%); split: -0.16%, +1.48%
InvThroughput: 462960 -> 461922 (-0.22%); split: -0.42%, +0.19%

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17697>

aco: fix LdsBranchVmemWARHazard with 2+ branch chains

For example, "DS -> branch -> VMEM -> branch -> DS".

fossil-db (navi10):
Totals from 639 (0.40% of 161220) affected shaders:
Instrs: 629090 -> 628254 (-0.13%); split: -0.19%, +0.06%
CodeSize: 3410164 -> 3406748 (-0.10%); split: -0.14%, +0.04%
Latency: 7834755 -> 7821011 (-0.18%); split: -0.70%, +0.52%
InvThroughput: 1369698 -> 1374495 (+0.35%); split: -0.12%, +0.47%

A lot of the fossil-db changes are noise.
threekingdoms.8db138826c386a62.1.foz/0b222ed175eebad0 is an example of a
shader that actually has this issue.

Signed-off-by: Rhys Perry <pendingchaos02@gmail.com>
Reviewed-by: Timur Kristóf <timur.kristof@gmail.com>
Fixes: c037ba1bb7a ("aco/gfx10: Mitigate LdsBranchVmemWARHazard.")
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17697>

gallium/u_threaded: buffer subdata merging (v2)

In a scenario where a sequence of calls happens like:
* subdata(buffer_a, offset=0, size=64)
* subdata(buffer_a, offset=64, size=64)
* subdata(buffer_a, offset=128, size=64)
* subdata(buffer_a, offset=192, size=64)

and the buffer can't be directly mapped (e.g., because it has bindings), the
subdata calls will now be merged together into one larger subdata call.

This achieves a 3x perf gain in
KHR-GL46.CommonBugs.CommonBug_SparseBuffersWithCopyOps on radeonsi

Before:
real    0m1,923s
user    0m1,017s
sys     0m0,051s

After:
real    0m0,686s
user    0m0,502s
sys     0m0,071s

Reviewed-By: Mike Blumenkrantz <michael.blumenkrantz@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17741>

tu: Flush depth on depth img transition from undef layout

Same logic as in tu_subpass_barrier.

Signed-off-by: Danylo Piliaiev <dpiliaiev@igalia.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17911>

ac/nir/cull: Fix typo in bounding box culling.

Bounding box culling is only viable when the W of all
vertices are positive. Always accept triangles whose any
W is negative.

Fixes: 0d527bb1aa720bf8e5735afdf8e9b70772e1ec23
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/7018
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Marek Olšák <marek.olsak@amd.com>
Reviewed-by: Qiang Yu <yuq825@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17929>

radv: use ref counting for VS prologs and PS epilogs

With GPL, it will be possible to create VS prologs and PS epilogs
from libraries, so reference counting is useful here too.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17628>

radv: rework shaders ref counting

Introduce helpers like for descriptor set layouts. This will also
help graphics pipeline libraries.

Signed-off-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Reviewed-By: Tatsuyuki Ishi <ishitatsuyuki@gmail.com>
Reviewed-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17628>

ci/radeonsi: Add zork jobs and rules

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/radeonsi: skqp: Add fail test files for raven

Lots of models are missing.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/freedreno: skqp: run with new tests files

Settings as flakes tests that passed in the exhaustive run, to keep the
same state as it was before

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Add gitlab sections for uncluttering

skqp output is verbose, as we are running multiple backends at the same
job, normally the trace will surpass the Gitlab UI line limit.

This commit wraps every skqp execution in a Gitlab section and removes
some `set -xtrace` from skqp-runner.sh for a cleaner output.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Remove .baremetal-skqp-test in favor of .skqp-test

Both hidden jobs has the same content, let's reuse it.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Supress irrelevant shellcheck warnings

To fix some warnings, one should write a much complex bash code, such as
SC2086, so prefer to be simple and functional.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Put generated tests files in artifacts

Showing the resulting test file can help the developer to debug skqp
runs by coping this file locally.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Use SKQP_BIN_DIR instead of hardcoded /skqp dir

This will make skqp-runner.sh more generic, making it easier to test
locally.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Add support for commenting tests files

The files are now separated in three: crashes, fails and flakes.

They should be located inside $INSTALL folder at:

- $GPU_VERSION_$SKQP_BACKEND_rendertests-$MODE.txt
- $GPU_VERSION_unittests-$MODE.txt

Where:

- $MODES can be crashes, fails, and flakes
- $SKQP_BACKEND can be gl, gles and vk

crashes and flakes removes tests from skqp, so they will not be run.
As skqp does not have support for flaky test detection, let's not run
them.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Build list_gpu_unit_tests and list_gms

These binaries are used to generate a list of tests that can be run in a
target device and are useful for testing new devices

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Fix Nima-Cpp fetching error

Nima-Cpp is not available anymore inside googlesource, revert to github
one

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Fix paths in skqp-runner

Default results directory was fixed via $PWD variable, but it is safer
to use the same as init-stage2.sh uses: $CI_PROJECT_DIR to indicate the
results folder.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Show reports on crashes

Some skqp tests may crash the entire job run, assure that the reports
will be showed to the user after the test started to run.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

ci/skqp: Add an option to run all tests

When the skqp is introduced to a new driver, the best practice is to
run all available tests from skqp and classifying the
failing/crashing/flaking ones.
The default behavior of skqp is to run the tests from the commit where
the skqp built, which may not be adequate for the target driver.

Signed-off-by: Guilherme Gallo <guilherme.gallo@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17835>

radv: ignore out-of-order rasterization if stencil write mask is dynamic

This might break out-of-order rasterization on GFX8-GFX9 because it
relies on the stencil write mask which can be dynamic.

Found by inspection.

Cc: mesa-stable
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/17673>

Revert "nir: Preserve offsets in lower_io_to_scalar_early"

This reverts commit 96fa23bca5ac88e0cd2dd0c45fdef71b2afe888d.

The correct fix to the problem was a1bc1523408a3, making this
change obsolete as the pass skips any vars marked with
always_active_io. There was no real advantage to allowing these
vars to be split because they can't be removed anyway. Also there
is no way to split varying arrays gracefully here due to the xfb
layout rules, and this change didn't handle arrays at all.

Removing this obsolete code also fixes an assert in the new CTS
test KHR-Single-GL45.enhanced_layouts.xfb_all_stages. The test
was legally adding xfb offsets to all vertex stages but since
we only mark the varyings in the final vertex stage with the
always_active_io flag the other stages were correctly lowering
to scalars but when an array with an offset hit this code it
asserted since it couldn't handle it.

Acked-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Fixes: a1bc1523408a3 ("spirv: mark variables decorated with XfbBuffer as always active")
Closes: https://gitlab.freedesktop.org/mesa/mesa/-/issues/6928
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17878>

agx: Only emit the used components of gl_FragCoord

In case a shader only use gl_FragCoord.xy, this avoids wasting
coefficient registers for gl_FragCoord.zw which should be a small
optimization. It's also less work for DCE but I'm less worried about
that.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Remove p_extract

It's now unused. We didn't have coalescing for it anyway, splits are the
preferred alternative.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Handle type-changing splits

If we want to break down a 64-bit value into its 32-bit halves, we want
to be able to use a split for this:

   lo, hi = split long

Extend the RA to handle this case.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Stop using broken idiv lowering

It is, as the name suggests, broken. Instruction count goes from 50->53
on the shader in
dEQP-GLES2.functional.shaders.operator.binary_operator.div.highp_int_fragment.
I'm happy to eat that cost in exchange for correct results!

There are lots more low-hanging opportunities for optimizations to that
shader:

- fuse double icmpsel for the b2i32(cmp) sequences
- promoting big immediates to uniforms
- fusing integer multiply+add

But for now this is acceptable and anyway I'm doing this on "fix broken
NIR lowering" time and not Asahi time.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Implement nir_op_umul_high

This is crucial to the efficiency of the accurate idiv path.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Extract umul_high implementation

We can implement umul_high (for both 16-bit and 32-bit types)
efficiently by multiplying in the next larger type size and extracting
the upper word. We already have such an implementation (for instancing).
Extract it so we can use it for emit_alu too.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Assert that registers are naturally aligned

This seems to be an architectural constraint. Ensure that RA satisfies
it, because otherwise we're left with mysterious fails.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Align 64-bit register pairs

This seems to be necessary for correct operation.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Lower more ALU operations

Noticed while switching idiv lowerings. We could do better on some of these.

Signed-off-by: Alyssa Rosenzweig <alyssa@collabora.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Implement noperspective interpolation

We need to get a matching coefficient register and change the encoding
of the iter instruction slightly, but otherwise this is normal.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Use split instead of extract for ldcf

For more uniform handling in the RA. This gets rid of the extra moves
with flat shading.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Rename varying load instructions

Unlike Mali (where I borrowed the old names from), these are not loads
in the memory sense. They are simply register loads and arithmetic.
Rename accordingly, using PowerVR names and public Apple names as a
guide.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Model perspective coefficient reg in the IR

For perspective-correct interpolation, the W coefficient register is
needed. Instead of hardcoding this to cf0 and special casing, model this
in the IR and let the general handling kick in.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

agx: Add AGX_MESA_DEBUG=noopt option

To disable the optimizer. Trying to root cause a Neverball bug, this
gives one less thing to worry.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>

asahi,agx: Rewrite varying linking

Instead of using driver_location magic and hoping things work, make the
linkage between vertex and fragment shaders explicit. Thanks to the
coefficient register mechanism reverse-engineered and documented earlier
in this series, this does not require any shader keys to support
separable shaders. It just requires that we regenerate the coefficient
register binding tables at draw time, based on the varying layouts
decided by the compiler independently for the VS and FS. This is more
robust in the face of separate shaders.

This also gets us glProvokingVertex() support without shader keys.

After that, we don't need any of the remapping prepasses. For fragment
shaders, any old mapping will do, so we can assign coefficient registers
as we go (based on what the program actually uses, not nir_variable
information that might be stale by this point). We do want to cache
coefficient registers, particularly for fragcoord.w which is used for
perspective interpolation everywhere.

Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/17198>