Below are the list of all the individual changes, which are a good order
in which to review this CL.
Core changes:
- Change the tracking in the frontend to produce a synchronization
scope per dispatch instead of per compute pass. Some bindgroups might
not be part of any synchronization scopes so we also track all the
referenced resources on the side so they can be checked during
Queue::Submit validation.
- Fix clearing in the GL and Metal backends to use the per-dispatch
synchronization scopes.
- Fix the Vulkan backend to use the per dispatch synchronization scopes
to produce the correct pipeline barriers. This allows the removal of
previous logic that was subtly incorrect for Indirect buffer. This
allows the merging of the Compute and Render DescriptorSetTracker into
a single small helper class.
- D3D12 changes are similar to Vulkan, but the simplification is just a
the suppression of a branch with a lot of code in
BindGroupStateTracker.
Test changes:
- Fixup all the ResourceUsageTracking tests to follow the WebGPU spec
for synchronization scopes (fixing a lot of TODOs).
- Add additional tests checking that Indirect buffers are not allowed
to be used as a writeable storage in the same synchronization scope.
- Add tests for Queue::Submit validation correctly taking into account
resources that are bound but unused in compute passes.
- Add an end2end test for using a buffer as Indirect and Storage at the
same time in a DispatchIndirect, which would previously produce
incorrect barriers in the Vulkan and D3D12 backends.
Other small changes (that I was to lazy to put in a different CL):
- Add the utils::MakePipelineLayout helper function.
- Fix Indirect not being in the list of readonly buffer usages (caught
by a test added in this CL).
Bug: dawn:632
Change-Id: I77263c3535a4ba995faccbf26255da9a2f6ed3b5
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/49887
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Stephen White <senorblanco@chromium.org>
This change is a preparation for making the compute pass track the
synchronization scope usages per dispatch instead of for the whole pass.
This CL just separates the Compute and RenderPassResourceUsage types.
This requires making the difference between
SyncScope/ComputePass/RenderPass ResourceUsageTracker instead of having
a single combined tracker.
This change also duplicates SetBindGroup by removing the common handling
in ProgrammablePassEncoder and putting it in ComputePassEncoder and
RenderEncoderBase. This is necessary because the UsageTracker types are
now split, but it will also help have different handling of SetBindGroup
for compute and render in follow-up CLs.
There are no functional changes.
Bug: dawn:632
Change-Id: I482c04483d8b734fb10e44e717071eedcff2f15f
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/49884
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
This change is a preparation for making the compute pass track the
synchronization scope usages per dispatch instead of for the whole pass.
- Split the tracking of render and compute passes usages.
- Rename PassTextureUsage to TextureSubresourceUsage since is it not
per-pass
- Add SyncScopeResourceUsage as a subclass of PassResourceUsage to
start modifying some of the code to work with synchronization scopes
(even if syncscope == pass at the moment).
There are no functional changes.
Bug: dawn:632
Change-Id: Ieeb6d70a44dc1c726f26989eebcd87e63e732785
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/49883
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
We do query resetting for each query commands outside render pass,
whether they're rewritten or not, so no longer need to track their
availability on command encoder. The availability on query set is enough
for resolving sparse queries.
But we still need to track query availability on render pass for query
rewrite checking and query resetting per render pass. Because reset
command must be called outside render pass, we need to reset them
together before the beginning render pass based that. Add availability
tracking on pass resource usage tracker (we only need it on render pass
) to facilitate use it in Vulkan backend.
Bug: dawn:434
Change-Id: Ie1b413ff54f62f3b84fe612e4abe45872c387e81
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/45440
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Hao Li <hao.x.li@intel.com>
The global resource usage was used for two things:
- Validating that the texture had the required usage, but this was
removed in a previous commit in favor of checking the texture's usage at
the encoding entrypoint.
- Skipping laz-clearing of the texture if it was used as
RenderAttachment. This was incorrect and would skip clearing of all of
the texture's subresource as long as one of them was used as
RenderAttachment.
This commit make PassTextureUsage exactly a
SubresourceStorage<TextureUsage> and fixes the logic for skipping
the clearing or RenderAttachment. It also adds a regression test for the
lazy clearing fix.
Bug: dawn:635
Change-Id: I5d984febb3e5a5f9ae15b632cac68e294555c4e6
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/38382
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Stephen White <senorblanco@chromium.org>
Auto-Submit: Corentin Wallez <cwallez@chromium.org>
Using this container is a small performance regression in the simple
cases where all subresources are the same, or when the texture has few
subrsources. However it give better performance in the hard subresource
tracking cases of textures with many subresources.
Using SubresourceStorage also makes it easier to work with compressed
storage since the compression is mostly transparent. It reduces code
duplication and prevent bugs from appearing when a developer would
forget to handle compression.
This fixes a state tracking issue in ValidatePassResourceUsage where the
function didn't correctly handle the case where the PassResourceUsage
was compressed.
Also removes the unused vulkan::Texture::TransitionFullUsage.
Also makes SubresourceStorage<T> only require operator== on T and not
operator !=.
Also fixes the texture format's aspect being used to create pipeline
barriers instead of the range's aspects.
Bug: dawn:441
Change-Id: I234b8191f39a09b541c1c63a60cccd6cee970550
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/37706
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Auto-Submit: Corentin Wallez <cwallez@chromium.org>
In a render/compute pass, if a texture is used as the same usages
across its subresources, and it is also used as different usages
across its subresources. We should mark that it is not used as the
same usages across its subresources for this pass, and we need to
visit every single subresource to track its usages transition.
Bug: dawn:462
Change-Id: Ide0770db62a3b7b85664757b21eb3280a0bb646c
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/23300
Commit-Queue: Yunchao He <yunchao.he@intel.com>
Reviewed-by: Austin Eng <enga@chromium.org>
If a texture follow this rule:
* old usages of all subresources are the same
* new usages of all subresources are the same
Then we can use one barrier to transit states for all subresources.
We don't need to use one barrier per each subresource.
This patch adds a flag at front-end to capture many (but not all)
situations which follow the rule above, then we don't need to loop
over every subresource at Vulkan backend.
Bug: dawn:441
Change-Id: I6124ee2cb09c9142fefd8f057dc6d2659301e2d4
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/22702
Commit-Queue: Yunchao He <yunchao.he@intel.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
This patch also add validation tests for texture subresource tracking
for render pass. Resource usage tracking for compute is per each
dispatch() call, I will add it in next patch.
BUG=dawn:157
Change-Id: I6c4b932e317d66521fa428311e727876d0adf4ea
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/17661
Commit-Queue: Yunchao He <yunchao.he@intel.com>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
This patch enables one texture to be used as both write-only storage
texture and sampled texture in one compute pass.
Note that while we don't need to check the usage scope of a texture in
the whole compute pass scope, we still need to verify one texture cannot
be bound to multiple bindings that are used in one dispatch at the same
time. This check will be added in the following patches.
This patch also adds tests to ensure a texture can be used as the
following binding types in one compute pass:
- read-only storage + sampled
- write-only storage + sampled
- read-only storage + write-only storage
BUG=dawn:267
TEST=dawn_unittests
Change-Id: Ibff2b005a5269a0bfa254e0417de4920758add39
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/20120
Commit-Queue: Jiawei Shao <jiawei.shao@intel.com>
Reviewed-by: Austin Eng <enga@chromium.org>
This patch factors resource usage tracking so it is separate from
command validation, allowing the bulk of command validation to be
completely skipped.
In DrawCallPerfRun/Vulkan_DynamicPipeline_DynamicBindGroup, disabling
validation cuts roughly 74 nanoseconds (20%) of CPU time, per draw.
In DrawCallPerfRun/Vulkan_DynamicBindGroup, disabling validation
cuts roughly 35 nanoseconds (17%) of CPU time, per draw.
In DrawCallPerfRun/Vulkan_MultipleBindGroups, disabling validation
cuts roughly 45 nanoseconds (14%) of CPU time, per draw.
Bug: dawn:271
Change-Id: I517b85840ba18c6a554b83f34a1d0aef1a8c56a6
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/13520
Commit-Queue: Austin Eng <enga@chromium.org>
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
This is to match the naming convention of WebGPU's WebIDL and webgpu.h
BUG=dawn:22
Change-Id: Ia91c5a018403e6a72eb0311b5f1a072d102282a2
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/10461
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
Reviewed-by: Austin Eng <enga@chromium.org>
Likewise "presented" textures won't be available for use in submits so
this sets up state-tracking and validation for textures too.
Command buffer resource usage is now stored in the frontend instead of
done per-backend because it is used to validate resources are allowed in
the submits.
Also adds a test.
BUG=dawn:9
Change-Id: I0537c5113bb33a089509b4f2af4ddf4eff8051ea
Reviewed-on: https://dawn-review.googlesource.com/c/2142
Reviewed-by: Corentin Wallez <cwallez@chromium.org>
Commit-Queue: Corentin Wallez <cwallez@chromium.org>
The dawn.h and dawncpp.h structure definitions references dawnFoo or
dawn::Foo respectively when it should reference dawn_native::FooBase* in
dawn_native. Autogenerate files to declare the dawn_native version of
the structs and change the ProcTable generation to use it instead.
This is important to make libdawn_native a shared library because
currently it was depending on dawncpp's definition of .Get().
