I added the forcing of the "loop" attribute to all loops to address FXC
failing on uniformity errors related to gradients in loops. Since then,
Tint now implements UA and it recently became an error, so we no longer
need this hack. As a result, FXC is now better able to cope with loops
that it determines executes 0 times.
Most e2e tests are affected because so many use loops, but 27 tests that
were previously failing are now passing with this change:
tint/bug/tint/1538.wgsl.expected.fxc.hlsl
tint/bug/tint/1604.wgsl.expected.fxc.hlsl
tint/bug/tint/1605.wgsl.expected.fxc.hlsl
tint/unittest/reader/spirv/SpvParserCFGTest_ClassifyCFGEdges_LoopBreak_FromLoopHeader_SingleBlockLoop_TrueBranch.spvasm.expected.fxc.hlsl
tint/unittest/reader/spirv/SpvParserCFGTest_ComputeBlockOrder_Loop_HeaderHasBreakUnless.spvasm.expected.fxc.hlsl
tint/unittest/reader/spirv/SpvParserCFGTest_EmitBody_IfSelection_TrueBranch_LoopBreak.spvasm.expected.fxc.hlsl
tint/unittest/reader/spirv/SpvParserCFGTest_FindIfSelectionInternalHeaders_TrueBranch_LoopBreak_Ok.spvasm.expected.fxc.hlsl
tint/unittest/reader/spirv/SpvParserFunctionVarTest_EmitStatement_Phi_MultiBlockLoopIndex.spvasm.expected.fxc.hlsl
tint/unittest/reader/spirv/SpvParserFunctionVarTest_EmitStatement_Phi_SingleBlockLoopIndex.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/cov-dead-code-unreachable-merge/0-opt.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/cov-dead-code-unreachable-merge/0-opt.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/similar-nested-ifs/0-opt.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/similar-nested-ifs/0-opt.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/spv-load-from-frag-color/1.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/spv-load-from-frag-color/1.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-false-if-discard-loop/0.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-false-if-discard-loop/0.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-fragcoord-less-than-zero/0.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-fragcoord-less-than-zero/0.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-fragcoord-less-than-zero/1.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-fragcoord-less-than-zero/1.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-with-loop-read-write-global/0-opt.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-with-loop-read-write-global/0-opt.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-with-loop-read-write-global/1.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/stable-binarysearch-tree-with-loop-read-write-global/1.wgsl.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/write-red-after-search/0-opt.spvasm.expected.fxc.hlsl
tint/vk-gl-cts/graphicsfuzz/write-red-after-search/0-opt.wgsl.expected.fxc.hlsl
Bug: tint:1522
Bug: tint:1538
Bug: tint:1604
Bug: tint:1605
Change-Id: I530b846b6b8df122ab351ff7b85d3e1c9ac11526
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/104121
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Antonio Maiorano <amaiorano@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
This CL refactor the end-to-end test suit ComputeLayoutMemoryBufferTests
and add tests for non-struct-member scalar, vector, matrix, and array of
vectors and matrices types. This test suit is also intend to test f16
buffer read/write after it is implemented.
Bug: tint:1673
Change-Id: Iea4d3f70897d196ea00e3a3e0189a0372afe0382
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/102800
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Zhaoming Jiang <zhaoming.jiang@intel.com>
Adds suppressions for multiple failures in the end2end tests that are
showing up on Pixel devices I'm able to test. Majority (800+) affect
the devices with Qualcomm GPUs (Pixel 4, Pixel 2, etc) and 8 affect the
newer Tensor devices (Pixel 6).
Bug: dawn:1549
Bug: dawn:1550
Change-Id: Ia598734a1752e5f086e4e79c96a799156d84e448
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103940
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Brandon Jones <bajones@chromium.org>
Calling ID3D12SharingContract::Present issues GPU work on the command
queue which needs to be synchronized with resource deallocation. This
seems to work now perhaps due to the keyed mutex semantics keeping the
D3D11 texture alive for longer than necessary. With fences, this missing
synchronization causes the validation layers to complain about early
deallocation of the ID3D12Resource.
Moving the Present to SynchronizeImportTextureBeforeUse ensures that it
happens before NextSerial and hence the signal fence that's recorded
will include any GPU work issued by Present. Also, resource deallocation
will happen after this work. However, this has the side-effect of PIX
seeing more frames, once per ExecuteCommandLists, but it could be argued
that's more accurate and useful.
Bug: dawn:1544
Change-Id: I1b417049045a812837f67072d7f09ac47bc18125
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103841
Reviewed-by: Austin Eng <enga@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Sunny Sachanandani <sunnyps@chromium.org>
Currently deepest element only returns the type if it was a scalar
or the element of a vector, array or matrix. Otherwise it would
return `nullptr`. There are cases where we want to get the deepest
type which is a struct or some other type.
This Cl updates ElementOf to return the type instead of nullptr for
types which previously returned nullptr.
Change-Id: I7963d4ce55d2e2b1a537a7533fa332813eed035c
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103900
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: Dan Sinclair <dsinclair@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Unused aspects of depth-stencil attachments that are tagged as read-only
used to leak that read-only state to backends, even if the validation
made it seems like they always match (they only need to match if the
texture has both aspects). This confused backends like Vulkan which
checked for depthReadOnly || stencilReadOnly to choose between code
paths.
Instead reyify the depthStencilAttachement descriptor in the frontend to
protect against garbage values being passed for aspects that aren't
present in the texture.
Adds a regression test, with the caveat that a failure is only shown by
having the VVL output and error in stderr due to an unrelated issue.
Fixed: dawn:1512
Change-Id: I35d5581e46909b7f41ff4c7553d60c6ac844a56b
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/101121
Reviewed-by: Loko Kung <lokokung@google.com>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
T2T copies for depth-stencil formats where done one by one. Because
the two per-aspect copies where submitted with no barriers in between
them, incorrect synchronization could occur, making the end state of the
destination texture incorrect.
Fix this by using the combined aspects of the texture to perform copies,
such that depth stencil are copied in a single command instead of two
commands.
Unfortunately the VVLs don't catch this issue, but the reporter of the
issue confirmed that this commit fix the dawn_end2end_tests failures
they were seeing.
Fixed: dawn:1514
Change-Id: I2e1c5f8d9aabeb0119364d26c9d66d0763cfadcf
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103421
Kokoro: Kokoro <noreply+kokoro@google.com>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Auto-Submit: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Loko Kung <lokokung@google.com>
At the moment the computations to decide whether aspects should be
combined are executed on every call related to aspect in TextureVk.
These computations never change and can be computed once at the creation
of TextureVk and reused at runtime.
This is meant to be a noop change as a slight rework prior to fixing
depth-stencil T2T copies no using the combined aspects.
Bug: dawn:1514
Change-Id: I1177cdcf42d072bb2bc2c3a2f149dc480fe79f2f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103420
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Loko Kung <lokokung@google.com>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reverts to the code flow before fences were implemented. NextSerial is
now the reponsibility of the caller of ExecutePendingCommandContext like
it was before. We now use GetPendingCommandSerial to store the signal
fence value instead of GetLastSubmittedCommandSerial and check that the
signal fence value was submitted in EndAccess.
Bug: dawn:576
Change-Id: I616840a0932ec17f77fcab38058773006dfae32f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103501
Commit-Queue: Sunny Sachanandani <sunnyps@chromium.org>
Auto-Submit: Sunny Sachanandani <sunnyps@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
This patch adds DawnAdapterPropertiesPowerPreferenceDescriptor for
querying adapter power preference which is useful to distinguish
different logical adapters created on same physical device but with
different power preferences.
Bug: dawn:1516
Test: dawn_unittests
Change-Id: I12ed6e370f8b57c860520154565765f0ee894831
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/102780
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Jiawei Shao <jiawei.shao@intel.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
- Cleanup is necessary because otherwise encoded render commands may be
leaked if the validation encoding fails. (The leaked render commands
can then trigger an assert in ~Device::Cache because the commands can
hold a ref to an AttachmentState that was not destroyed, and hence
still be in the device cache.
- Added explicit check in EncoderIndirectDrawValidationCommands for
device 'alive-ness' since it may create new objects later and hit the
same error later on anyways.
- Added regression test.
Fixed: chromium:1365011
Change-Id: I342479a4227fc43d82ea35f662d049e6db2b1740
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103340
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Loko Kung <lokokung@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
A `@workgroup_size()` value must be a constant or override expression.
There's nothing specific here about literals or variable expressions.
Remove the semantic tracking of override variables, as these can be override expressions.
The backends will require the `SubstituteOverride` transform to be run, so gut the workgroup_size override handling from the backends.
Bug: tint:1633
Change-Id: Ib3ff843fc64a3595d49223c661b4d58130c0ab30
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/100142
Commit-Queue: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Dan Sinclair <dsinclair@chromium.org>
Previously when moving around directories for generated files, Dawn ran
into an issue where stale files where #included instead of the new ones,
causing compilation failures. To get around this a
remove_stale_autogen_files mechanism was added that scans the gen/
directory for files not in an allow-list of directories.
This mechanism is now causing problems for bringing up Dawn standalone
tests on Android as these test also generate files in Dawn's gen/
directories, and their files get deleted by remove_stale_autogen_files.
We are not foresseing any additional shuffling of directories and it's
safe to expect that all stale files have been removed from CI builder
caches at this time. So remove_stale_autogen_files can go. This is what
this CL does.
Fixed: dawn:1543
Change-Id: I7dbf1eae6c55b7659f3837b6d4a565052001ce57
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103040
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Exporting an exportable VkSemaphore doesn't implicitly destroy the
VkSemaphore object. So instead of Detach()ing the VkSemaphore when it is
first consumed, just let it go out of the scope and be destroyed with
RAII. This also fixes the RAII by not destroying the VkSemaphore
immediately and instead wait until it becomes unused.
Found by running dawn_end2end_tests with the VVLs.
Bug: chromium:1258986
Change-Id: I858839b3094eee0f575c07a8f18504680afb53e3
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103024
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Loko Kung <lokokung@google.com>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
While trying to debug why VVL failures don't cause test failures I
reworked this code a little bit. There is not CL that fixes the
behavior, but the code is marginally better with less indentation so
here's a CL to check that in.
Bug: None
Change-Id: I6fc460c4b4b7959ae405219615a03230bfb9847a
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/103022
Reviewed-by: Austin Eng <enga@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
With the change to remove support for `Override` in the various
backends, it is now possible for the fuzzers to send invalid programs
through to the generators by creating overrides.
This CL adds the `SubstituteOverride` transform into the fuzzers and
defaults any non-initialized override to 0. The transform is run
separate from the other transforms used by the fuzzers as the fuzzers
don't have to add transforms, this makes sure the substitution always
happens, regardless of other transform configuration.
Bug: chromium:1362815
Change-Id: I3c57128d24c5613079a62309f5d5edefa28e8413
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/102840
Commit-Queue: Dan Sinclair <dsinclair@chromium.org>
Reviewed-by: Ryan Harrison <rharrison@chromium.org>
Kokoro: Kokoro <noreply+kokoro@google.com>
Shuffle the transform orders to ensure that these are embedded in a structure before running the Std140 transform.
Add more end-to-end tests for these.
As pointed out in tint:1673, arrays of matrices are not correctly decomposed by the Std140 transform.
This will be addressed by a later change.
Bug: tint:1673
Change-Id: I47c93e458ff48578922d576819792e8ed3a5723c
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/102541
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Zhaoming Jiang <zhaoming.jiang@intel.com>
Commit-Queue: Ben Clayton <bclayton@google.com>
Reviewed-by: David Neto <dneto@google.com>
We already compute the "first" and "last" basic block that
uses a value, so we could know when to hoist a value into
a var declaration. You have to do this sometimes to make
sure all uses are in scope of the declaration.
Until now we tracked Phis with an entirely different mechanism.
But there are cases which broke down. That's what happens
in crbug.com/tint/1649.
Additionally, GraphicsFuzz cases generarte similar weirdness.
Also, be more careful about ensuring that the assignments
generated to feed phis behave as if they occur in parallel.
Within a single batch of such assignments, generate and
use intermediate let-declarations for phis that that batch
will overwrite.
Also, unwrap-references when rectifying the signedness of
binary operators.
Skip tests that fail due to crbug.comt/tint/98:
test/tint/bug/tint/749.spvasm.*
Fixed: tint:1649
Change-Id: I7314c351b74a10bfa9a18011f3d80a520568011c
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/101220
Auto-Submit: David Neto <dneto@google.com>
Commit-Queue: David Neto <dneto@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
