tree d7f87b485df2ecce51f0e673bbf753cb280ab5e6
parent 0aa192ef1bb53cd65ee1ccdfb1bf23e98c2e6e65
author mmaxfield@apple.com <mmaxfield@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc> 1560406013 +0000
committer mmaxfield@apple.com <mmaxfield@apple.com@268f45cc-cd09-0410-ab3c-d52691b4dbfc> 1560406013 +0000

[WHLSL] Hook up compute
https://bugs.webkit.org/show_bug.cgi?id=198644

Reviewed by Saam Barati.

Source/WebCore:

This patch hooks up compute shaders in exactly the same way that vertex and fragment shaders
are hooked up. I've modified the two patchs (compute and rendering) to be almost exactly the
same code.

This patch also adds support for the WHLSL compiler to determine what the numthreads()
attribute in the shader says so that it can be hooked up to Metal's threads-per-threadgroup
argument in the dispatch call. There is some logic to make sure that there aren't two
numthreads() attributes on the same compute shader.

It also adds a little bit of type renaming. For built-in variables, sometimes Metal's type
doesn't always match WHLSL's (and HLSL's type). For example, in WHLSL and HLSL,        SV_DispatchThreadID variables have to be a float3, but in Metal, they are a uint3.
Therefore, I've added a little bit of code during each entry point's pack and unpack stages
to handle this type conversion.

Test: webgpu/whlsl-compute.html

* Modules/webgpu/WHLSL/Metal/WHLSLEntryPointScaffolding.cpp:
(WebCore::WHLSL::Metal::internalTypeForSemantic): Determine which Metal type corresponds to
each built-in variable.
(WebCore::WHLSL::Metal::EntryPointScaffolding::builtInsSignature): Perform the type
conversion.
(WebCore::WHLSL::Metal::EntryPointScaffolding::unpackResourcesAndNamedBuiltIns): Ditto.
(WebCore::WHLSL::Metal::VertexEntryPointScaffolding::VertexEntryPointScaffolding): Ditto.
(WebCore::WHLSL::Metal::VertexEntryPointScaffolding::helperTypes): Ditto.
(WebCore::WHLSL::Metal::VertexEntryPointScaffolding::pack): Ditto.
(WebCore::WHLSL::Metal::FragmentEntryPointScaffolding::FragmentEntryPointScaffolding): Ditto.
(WebCore::WHLSL::Metal::FragmentEntryPointScaffolding::helperTypes): Ditto.
(WebCore::WHLSL::Metal::FragmentEntryPointScaffolding::pack): Ditto.
(WebCore::WHLSL::Metal::ComputeEntryPointScaffolding::signature): Ditto.
* Modules/webgpu/WHLSL/Metal/WHLSLEntryPointScaffolding.h:
* Modules/webgpu/WHLSL/WHLSLComputeDimensions.cpp: Added. Add a pass to determine whether
or not any entry point has duplicate numthreads() attribute, and to determine what the
appropriate numthreads() values should be for the current entry point.
(WebCore::WHLSL::ComputeDimensionsVisitor::ComputeDimensionsVisitor):
(WebCore::WHLSL::ComputeDimensionsVisitor::computeDimensions const):
(WebCore::WHLSL::computeDimensions):
* Modules/webgpu/WHLSL/WHLSLComputeDimensions.h: Copied from Source/WebCore/Modules/webgpu/WHLSL/WHLSLPrepare.h.
* Modules/webgpu/WHLSL/WHLSLGatherEntryPointItems.cpp:
(WebCore::WHLSL::gatherEntryPointItems): Compute shaders don't need to have a semantic for their return type.
* Modules/webgpu/WHLSL/WHLSLPrepare.cpp:
(WebCore::WHLSL::prepare): Run the computeDimensions() pass.
* Modules/webgpu/WHLSL/WHLSLPrepare.h:
* Modules/webgpu/WHLSL/WHLSLPropertyResolver.cpp: In a left-value propertyAccessExpression,
the index expression can be a right-value. Treat it as such.
(WebCore::WHLSL::LeftValueSimplifier::finishVisiting):
(WebCore::WHLSL::LeftValueSimplifier::visit):
* Modules/webgpu/WHLSL/WHLSLStandardLibrary.txt: We need support for multiplication (for a
test) and float3 for SV_DispatchThreadID.
* Sources.txt:
* SourcesCocoa.txt:
* WebCore.xcodeproj/project.pbxproj:
* platform/graphics/gpu/GPUComputePipeline.h: Associate a compute dimensions with a particular
compute pipeline. This is how Metal knows what values to use for a dispatch.
(WebCore::GPUComputePipeline::computeDimensions const):
* platform/graphics/gpu/cocoa/GPUComputePassEncoderMetal.mm: Use the saved compute dimensions.
(WebCore::GPUComputePassEncoder::dispatch):
* platform/graphics/gpu/cocoa/GPUComputePipelineMetal.mm: Make the code match GPURenderPipelineMetal.
(WebCore::trySetMetalFunctions):
(WebCore::trySetFunctions):
(WebCore::convertComputePipelineDescriptor):
(WebCore::tryCreateMTLComputePipelineState):
(WebCore::GPUComputePipeline::tryCreate):
(WebCore::GPUComputePipeline::GPUComputePipeline):
(WebCore::tryCreateMtlComputeFunction): Deleted.
* platform/graphics/gpu/cocoa/GPUPipelineMetalConvertLayout.cpp: Added. Moved shared helper
functions to a file where they can be accessed by multiple places.
(WebCore::convertShaderStageFlags):
(WebCore::convertBindingType):
(WebCore::convertLayout):
* platform/graphics/gpu/cocoa/GPUPipelineMetalConvertLayout.h: Copied from Source/WebCore/Modules/webgpu/WHLSL/WHLSLPrepare.h.
* platform/graphics/gpu/cocoa/GPURenderPipelineMetal.mm: Delete the functions that were moved to GPUPipelineMetalConvertLayout.
(WebCore::trySetFunctions):
(WebCore::tryCreateMtlRenderPipelineState):
(WebCore::convertShaderStageFlags): Deleted.
(WebCore::convertBindingType): Deleted.
(WebCore::convertLayout): Deleted.

LayoutTests:

This doesn't thoroughly test compute, but it's at least enough to unblock the WHLSL testing effort.

* webgpu/compute-squares-expected.txt: Deleted. Covered by webgpu/whlsl-compute.html.
* webgpu/compute-squares.html: Deleted. Ditto.
* webgpu/whlsl-compute-expected.txt: Added.
* webgpu/whlsl-compute.html: Added.

git-svn-id: http://svn.webkit.org/repository/webkit/trunk@246396 268f45cc-cd09-0410-ab3c-d52691b4dbfc