src/dawn/tests/perf_tests/ShaderRobustnessPerf.cpp - dawn - Git at Google

 // Copyright 2021 The Dawn Authors
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
 // You may obtain a copy of the License at
 //
 //     http://www.apache.org/licenses/LICENSE-2.0
 //
 // Unless required by applicable law or agreed to in writing, software
 // distributed under the License is distributed on an "AS IS" BASIS,
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.

 #include <string>
 #include <vector>

 #include "dawn/tests/perf_tests/DawnPerfTest.h"
 #include "dawn/utils/WGPUHelpers.h"

 namespace dawn {
 namespace {

 constexpr uint32_t kTileSize = 32u;

 const std::string& kMatMulFloatHeader = R"(
         struct Uniforms {
             dimAOuter : u32,
             dimInner : u32,
             dimBOuter : u32,
         }
         struct Matrix {
             numbers: array<f32>
         }

         @group(0) @binding(0) var<storage, read> firstMatrix : Matrix;
         @group(0) @binding(1) var<storage, read> secondMatrix : Matrix;
         @group(0) @binding(2) var<storage, read_write> resultMatrix : Matrix;
         @group(0) @binding(3) var<uniform> uniforms : Uniforms;

         fn mm_readA(row : u32, col : u32) -> f32  {
             if (row < uniforms.dimAOuter && col < uniforms.dimInner)
             {
                 let result : f32 = firstMatrix.numbers[row * uniforms.dimInner + col];
                 return result;
             }
             return 0.;
         }

         fn mm_readB(row : u32, col : u32) -> f32 {
             if (row < uniforms.dimInner && col < uniforms.dimBOuter)
             {
                 let result : f32 = secondMatrix.numbers[row * uniforms.dimBOuter + col];
                 return result;
             }
             return 0.;
         }

         fn mm_write(row : u32, col : u32, value : f32) {
             if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)
             {
                 let index : u32 = col + row * uniforms.dimBOuter;
                 resultMatrix.numbers[index] = value;
             }
         }

         const RowPerThread : u32 = 4u;
         const ColPerThread : u32 = 4u;
         const TileAOuter : u32 = 32u;
         const TileBOuter : u32 = 32u;
         const TileInner : u32 = 32u;)";

 const std::string& kMatMulFloatSharedArray1D = R"(
         var<workgroup> mm_Asub : array<f32, 1024>;
         var<workgroup> mm_Bsub : array<f32, 1024>;)";
 const std::string& kMatMulFloatSharedArray2D = R"(
         var<workgroup> mm_Asub : array<array<f32, 32>, 32>;
         var<workgroup> mm_Bsub : array<array<f32, 32>, 32>;)";
 const std::string& kMatMulFloatBodyPart1 = R"(
         @compute @workgroup_size(8, 8, 1)
         fn main(@builtin(local_invocation_id) local_id : vec3u,
                 @builtin(global_invocation_id) global_id  : vec3u) {
             let tileRow : u32 = local_id.y * RowPerThread;
             let tileCol : u32 = local_id.x * ColPerThread;

             let globalRow : u32 = global_id.y * RowPerThread;
             let globalCol : u32 = global_id.x * ColPerThread;

             let numTiles : u32 = (uniforms.dimInner - 1u) / TileInner + 1u;

             var acc: array<f32, 16>;
             var ACached : f32;
             var BCached : array<f32, 4>;

             let ColPerThreadA : u32 = TileInner / 8u;
             let tileColA : u32 = local_id.x * ColPerThreadA;
             let RowPerThreadB : u32 = TileInner / 8u;
             let tileRowB : u32 = local_id.y * RowPerThreadB;

             // Loop over shared dimension.
             for (var t : u32 = 0u; t < numTiles; t = t + 1u) {
                 // Load one tile of A into local memory.
                 for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
                 for (var innerCol : u32 = 0u; innerCol < ColPerThreadA; innerCol = innerCol + 1u) {
                     let inputRow : u32 = tileRow + innerRow;
                     let inputCol : u32 = tileColA + innerCol;)";
 const std::string& kMatMulFloatBodyPart2Array1D = R"(
                     let index : u32 = inputRow * TileInner + inputCol;
                     mm_Asub[index] = mm_readA(globalRow + innerRow, t * TileInner + inputCol);
                 }
                 }
                 // Load one tile of B into local memory.
                 for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
                 for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
                     let inputRow : u32 = tileRowB + innerRow;
                     let inputCol : u32 = tileCol + innerCol;
                     let index : u32 = inputRow * TileBOuter + inputCol;

                     mm_Bsub[index] = mm_readB(t * TileInner + inputRow, globalCol + innerCol);;
                 }
                 }

                 workgroupBarrier();

                 // Compute acc values for a single thread.
                 for (var k : u32 = 0u; k < TileInner; k = k + 1u) {
                     for (var inner : u32 = 0u; inner < ColPerThread; inner = inner + 1u) {
                         BCached[inner] = mm_Bsub[k * TileBOuter + tileCol + inner];
                     }

                     for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
                         ACached = mm_Asub[(tileRow + innerRow) * TileInner + k];)";
 const std::string& kMatMulFloatBodyPart2Array2D = R"(
                     mm_Asub[inputRow][inputCol] = mm_readA(globalRow + innerRow, t * TileInner + inputCol);
                 }
                 }
                 // Load one tile of B into local memory.
                 for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
                 for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
                     let inputRow : u32 = tileRowB + innerRow;
                     let inputCol : u32 = tileCol + innerCol;

                     mm_Bsub[innerCol][inputCol] = mm_readB(t * TileInner + inputRow, globalCol + innerCol);;
                 }
                 }

                 workgroupBarrier();

                 // Compute acc values for a single thread.
                 for (var k : u32 = 0u; k < TileInner; k = k + 1u) {
                     for (var inner : u32 = 0u; inner < ColPerThread; inner = inner + 1u) {
                         BCached[inner] = mm_Bsub[k][tileCol + inner];
                     }

                     for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
                         ACached = mm_Asub[tileRow + innerRow][k];)";
 const std::string& kMatMulFloatBodyPart3 = R"(
                         for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
                             let index : u32 = innerRow * ColPerThread + innerCol;
                             acc[index] = acc[index] + ACached * BCached[innerCol];
                         }
                     }
                 }

                 workgroupBarrier();
             }

             for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
             for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
                 let index : u32 = innerRow * ColPerThread + innerCol;
                 mm_write(globalRow + innerRow,
                          globalCol + innerCol,
                          acc[index]);
             }
             }
         })";
 const std::string& kMatMulFloatOneDimensionalSharedArray =
     kMatMulFloatHeader + kMatMulFloatSharedArray1D + kMatMulFloatBodyPart1 +
     kMatMulFloatBodyPart2Array1D + kMatMulFloatBodyPart3;

 const std::string& kMatMulFloatTwoDimensionalSharedArray =
     kMatMulFloatHeader + kMatMulFloatSharedArray2D + kMatMulFloatBodyPart1 +
     kMatMulFloatBodyPart2Array2D + kMatMulFloatBodyPart3;

 // The vec4 version requires that dimInner and dimBOuter are divisible by 4.
 const std::string& kMatMulVec4Header = R"(
         struct Uniforms {
             dimAOuter : u32,
             dimInner : u32,
             dimBOuter : u32,
         }
         struct Matrix {
             numbers: array<vec4f>
         }

         @group(0) @binding(0) var<storage, read> firstMatrix : Matrix;
         @group(0) @binding(1) var<storage, read> secondMatrix : Matrix;
         @group(0) @binding(2) var<storage, read_write> resultMatrix : Matrix;
         @group(0) @binding(3) var<uniform> uniforms : Uniforms;

         fn mm_readA(row : u32, col : u32) -> vec4f  {
             if (row < uniforms.dimAOuter && col < uniforms.dimInner)
             {
                 let result : vec4f = firstMatrix.numbers[row * uniforms.dimInner / 4u + col];
                 return result;
             }
             return vec4f(0.0, 0.0, 0.0, 0.0);
         }

         fn mm_readB(row : u32, col : u32) -> vec4f {
             if (row < uniforms.dimInner && col < uniforms.dimBOuter)
             {
                 let result : vec4f = secondMatrix.numbers[row * uniforms.dimBOuter / 4u + col];
                 return result;
             }
             return vec4f(0.0, 0.0, 0.0, 0.0);
         }

         fn mm_write(row : u32, col : u32, value : vec4f) {
             if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)
             {
                 let index : u32 = col + row * uniforms.dimBOuter / 4u;
                 resultMatrix.numbers[index] = value;
             }
         }

         const RowPerThread : u32 = 4u;
         const ColPerThread : u32 = 4u;
         const TileOuter : u32 = 32u;
         const TileInner : u32 = 32u;)";
 const std::string& kMatMulVec4SharedArray1D = R"(
         var<workgroup> mm_Asub : array<vec4f, 256>;
         var<workgroup> mm_Bsub : array<vec4f, 256>;)";
 const std::string& kMatMulVec4SharedArray2D = R"(
         var<workgroup> mm_Asub : array<array<vec4f, 8>, 32>;
         var<workgroup> mm_Bsub : array<array<vec4f, 8>, 32>;)";
 const std::string& kMatMulVec4BodyPart1 = R"(
         @compute @workgroup_size(8, 8, 1)
         fn main(@builtin(local_invocation_id) local_id : vec3u,
                 @builtin(global_invocation_id) global_id  : vec3u) {
             let tileRow : u32 = local_id.y * RowPerThread;
             let tileCol : u32 = local_id.x;

             let globalRow : u32 = global_id.y * RowPerThread;
             let globalCol : u32 = global_id.x;

             let numTiles : u32 = (uniforms.dimInner - 1u) / TileInner + 1u;

             var acc: array<vec4f, 4>;
             var ACached : vec4f;
             var BCached : array<vec4f, 4>;

             var globalColA : u32 = tileCol;
             let RowPerThreadB : u32 = TileInner / 8u;
             let tileRowB : u32 = local_id.y * RowPerThreadB;

             // Loop over shared dimension.
             for (var t : u32 = 0u; t < numTiles; t = t + 1u) {
                 // Load one tile of A into local memory.
                 for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
                     let inputRow : u32 = tileRow + innerRow;
                     let inputCol : u32 = tileCol;)";
 const std::string& kMatMulVec4BodyPart2Array1D = R"(
                     let index : u32 = inputRow * TileInner / ColPerThread + inputCol;
                     mm_Asub[index] = mm_readA(globalRow + innerRow, globalColA);
                 }
                 globalColA = globalColA + TileInner / ColPerThread;

                 // Load one tile of B into local memory.
                 for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
                     let inputRow : u32 = tileRowB + innerRow;
                     let inputCol : u32 = tileCol;
                     let index : u32 = inputRow * TileOuter / ColPerThread + inputCol;
                     mm_Bsub[index] = mm_readB(t * TileInner + inputRow, globalCol);;
                 }

                 workgroupBarrier();

                 // Compute acc values for a single thread.
                 for (var k : u32 = 0u; k < TileInner / ColPerThread; k = k + 1u) {
                     BCached[0] = mm_Bsub[(k * ColPerThread) * (TileOuter / ColPerThread) + tileCol];
                     BCached[1] = mm_Bsub[(k * ColPerThread + 1u) * (TileOuter / ColPerThread) + tileCol];
                     BCached[2] = mm_Bsub[(k * ColPerThread + 2u) * (TileOuter / ColPerThread) + tileCol];
                     BCached[3] = mm_Bsub[(k * ColPerThread + 3u) * (TileOuter / ColPerThread) + tileCol];

                     for (var i : u32 = 0u; i < RowPerThread; i = i + 1u) {
                         ACached = mm_Asub[(tileRow + i) * (TileInner / ColPerThread) + k];)";
 const std::string& kMatMulVec4BodyPart2Array2D = R"(
                     mm_Asub[inputRow][inputCol] = mm_readA(globalRow + innerRow, globalColA);
                 }
                 globalColA = globalColA + TileInner / ColPerThread;

                 // Load one tile of B into local memory.
                 for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
                     let inputRow : u32 = tileRowB + innerRow;
                     let inputCol : u32 = tileCol;
                     mm_Bsub[inputRow][inputCol] = mm_readB(t * TileInner + inputRow, globalCol);;
                 }

                 workgroupBarrier();

                 // Compute acc values for a single thread.
                 for (var k : u32 = 0u; k < TileInner / ColPerThread; k = k + 1u) {
                     BCached[0] = mm_Bsub[k * ColPerThread][tileCol];
                     BCached[1] = mm_Bsub[k * ColPerThread + 1u][tileCol];
                     BCached[2] = mm_Bsub[k * ColPerThread + 2u][tileCol];
                     BCached[3] = mm_Bsub[k * ColPerThread + 3u][tileCol];

                     for (var i : u32 = 0u; i < RowPerThread; i = i + 1u) {
                         ACached = mm_Asub[tileRow + i][k];)";
 const std::string& kMatMulVec4BodyPart3 = R"(
                         acc[i] = BCached[0] * ACached.x + acc[i];
                         acc[i] = BCached[1] * ACached.y + acc[i];
                         acc[i] = BCached[2] * ACached.z + acc[i];
                         acc[i] = BCached[3] * ACached.w + acc[i];
                     }
                 }

                 workgroupBarrier();
             }

             for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
                 mm_write(globalRow + innerRow,
                          globalCol,
                          acc[innerRow]);
             }
         })";

 const std::string& kMatMulVec4OneDimensionalSharedArray =
     kMatMulVec4Header + kMatMulVec4SharedArray1D + kMatMulVec4BodyPart1 +
     kMatMulVec4BodyPart2Array1D + kMatMulVec4BodyPart3;

 const std::string& kMatMulVec4TwoDimensionalSharedArray =
     kMatMulVec4Header + kMatMulVec4SharedArray2D + kMatMulVec4BodyPart1 +
     kMatMulVec4BodyPart2Array2D + kMatMulVec4BodyPart3;

 constexpr unsigned int kNumIterations = 50;

 enum class MatMulMethod {
     MatMulFloatOneDimSharedArray,
     MatMulFloatTwoDimSharedArray,
     MatMulVec4OneDimSharedArray,
     MatMulVec4TwoDimSharedArray
 };

 std::ostream& operator<<(std::ostream& ostream, const MatMulMethod& matMulMethod) {
     switch (matMulMethod) {
         case MatMulMethod::MatMulFloatOneDimSharedArray:
             ostream << "MatMulFloatOneDimSharedArray";
             break;
         case MatMulMethod::MatMulFloatTwoDimSharedArray:
             ostream << "MatMulFloatTwoDimSharedArray";
             break;
         case MatMulMethod::MatMulVec4OneDimSharedArray:
             ostream << "MatMulVec4OneDimSharedArray";
             break;
         case MatMulMethod::MatMulVec4TwoDimSharedArray:
             ostream << "MatMulVec4TwoDimSharedArray";
             break;
     }
     return ostream;
 }

 using DimAOuter = uint32_t;
 using DimInner = uint32_t;
 using DimBOuter = uint32_t;
 DAWN_TEST_PARAM_STRUCT(ShaderRobustnessParams, MatMulMethod, DimAOuter, DimInner, DimBOuter);

 // Test the execution time of matrix multiplication (A [dimAOuter, dimInner] * B [dimInner,
 // dimBOuter]) on the GPU and see the difference between robustness on and off.
 class ShaderRobustnessPerf : public DawnPerfTestWithParams<ShaderRobustnessParams> {
   public:
     ShaderRobustnessPerf()
         : DawnPerfTestWithParams(kNumIterations, 1),
           mDimAOuter(GetParam().mDimAOuter),
           mDimInner(GetParam().mDimInner),
           mDimBOuter(GetParam().mDimBOuter) {}
     ~ShaderRobustnessPerf() override = default;

     void SetUp() override;

   private:
     void Step() override;

     wgpu::BindGroup mBindGroup;
     wgpu::ComputePipeline mPipeline;
     uint32_t mDimAOuter;
     uint32_t mDimInner;
     uint32_t mDimBOuter;
 };

 void ShaderRobustnessPerf::SetUp() {
     DawnPerfTestWithParams<ShaderRobustnessParams>::SetUp();

     const size_t dataASize = mDimAOuter * mDimInner;
     std::vector<float> dataA(dataASize);
     uint64_t byteASize = sizeof(float) * dataA.size();
     // It's ok to use all zeros to do the matrix multiplication for performance test.
     wgpu::Buffer bufA =
         utils::CreateBufferFromData(device, dataA.data(), byteASize, wgpu::BufferUsage::Storage);

     const size_t dataBSize = mDimInner * mDimBOuter;
     std::vector<float> dataB(dataBSize);
     uint64_t byteBSize = sizeof(float) * dataB.size();
     wgpu::Buffer bufB =
         utils::CreateBufferFromData(device, dataB.data(), byteBSize, wgpu::BufferUsage::Storage);

     uint64_t byteDstSize = sizeof(float) * mDimAOuter * mDimBOuter;
     wgpu::BufferDescriptor desc = {};
     desc.usage = wgpu::BufferUsage::Storage;
     desc.size = byteDstSize;
     wgpu::Buffer dst = device.CreateBuffer(&desc);

     uint32_t uniformData[] = {mDimAOuter, mDimInner, mDimBOuter};
     wgpu::Buffer uniformBuffer = utils::CreateBufferFromData(
         device, uniformData, sizeof(uniformData), wgpu::BufferUsage::Uniform);

     wgpu::ShaderModule module;
     switch (GetParam().mMatMulMethod) {
         case MatMulMethod::MatMulFloatOneDimSharedArray: {
             module =
                 utils::CreateShaderModule(device, kMatMulFloatOneDimensionalSharedArray.c_str());
             break;
         }

         case MatMulMethod::MatMulFloatTwoDimSharedArray: {
             module =
                 utils::CreateShaderModule(device, kMatMulFloatTwoDimensionalSharedArray.c_str());
             break;
         }

         case MatMulMethod::MatMulVec4OneDimSharedArray: {
             module =
                 utils::CreateShaderModule(device, kMatMulVec4OneDimensionalSharedArray.c_str());
             break;
         }

         case MatMulMethod::MatMulVec4TwoDimSharedArray: {
             module =
                 utils::CreateShaderModule(device, kMatMulVec4TwoDimensionalSharedArray.c_str());
             break;
         }
     }

     wgpu::ComputePipelineDescriptor csDesc;
     csDesc.compute.module = module;
     csDesc.compute.entryPoint = "main";
     mPipeline = device.CreateComputePipeline(&csDesc);

     mBindGroup = utils::MakeBindGroup(device, mPipeline.GetBindGroupLayout(0),
                                       {
                                           {0, bufA, 0, byteASize},
                                           {1, bufB, 0, byteBSize},
                                           {2, dst, 0, byteDstSize},
                                           {3, uniformBuffer, 0, sizeof(uniformData)},
                                       });
 }

 void ShaderRobustnessPerf::Step() {
     wgpu::CommandBuffer commands;
     {
         wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
         if (SupportsTimestampQuery()) {
             RecordBeginTimestamp(encoder);
         }
         wgpu::ComputePassEncoder pass = encoder.BeginComputePass();
         pass.SetPipeline(mPipeline);
         pass.SetBindGroup(0, mBindGroup);
         for (unsigned int i = 0; i < kNumIterations; ++i) {
             pass.DispatchWorkgroups(ceil(static_cast<float>(mDimBOuter) / float{kTileSize}),
                                     ceil(static_cast<float>(mDimAOuter) / float{kTileSize}), 1);
         }
         pass.End();
         if (SupportsTimestampQuery()) {
             RecordEndTimestampAndResolveQuerySet(encoder);
         }

         commands = encoder.Finish();
     }

     queue.Submit(1, &commands);

     if (SupportsTimestampQuery()) {
         ComputeGPUElapsedTime();
     }
 }

 TEST_P(ShaderRobustnessPerf, Run) {
     RunTest();
 }

 DAWN_INSTANTIATE_TEST_P(ShaderRobustnessPerf,
                         {D3D12Backend(), D3D12Backend({"disable_robustness"}, {}), MetalBackend(),
                          MetalBackend({"disable_robustness"}, {}), OpenGLBackend(),
                          OpenGLBackend({"disable_robustness"}, {}), VulkanBackend(),
                          VulkanBackend({"disable_robustness"}, {})},
                         {MatMulMethod::MatMulFloatOneDimSharedArray,
                          MatMulMethod::MatMulFloatTwoDimSharedArray,
                          MatMulMethod::MatMulVec4OneDimSharedArray,
                          MatMulMethod::MatMulVec4TwoDimSharedArray},
                         {512u},
                         {512u},
                         {512u});

 }  // anonymous namespace
 }  // namespace dawn
	// Copyright 2021 The Dawn Authors
	//
	// Licensed under the Apache License, Version 2.0 (the "License");
	// you may not use this file except in compliance with the License.
	// You may obtain a copy of the License at
	//
	// http://www.apache.org/licenses/LICENSE-2.0
	//
	// Unless required by applicable law or agreed to in writing, software
	// distributed under the License is distributed on an "AS IS" BASIS,
	// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
	// See the License for the specific language governing permissions and
	// limitations under the License.

	#include <string>
	#include <vector>

	#include "dawn/tests/perf_tests/DawnPerfTest.h"
	#include "dawn/utils/WGPUHelpers.h"

	namespace dawn {
	namespace {

	constexpr uint32_t kTileSize = 32u;

	const std::string& kMatMulFloatHeader = R"(
	struct Uniforms {
	dimAOuter : u32,
	dimInner : u32,
	dimBOuter : u32,
	}
	struct Matrix {
	numbers: array<f32>
	}

	@group(0) @binding(0) var<storage, read> firstMatrix : Matrix;
	@group(0) @binding(1) var<storage, read> secondMatrix : Matrix;
	@group(0) @binding(2) var<storage, read_write> resultMatrix : Matrix;
	@group(0) @binding(3) var<uniform> uniforms : Uniforms;

	fn mm_readA(row : u32, col : u32) -> f32 {
	if (row < uniforms.dimAOuter && col < uniforms.dimInner)
	{
	let result : f32 = firstMatrix.numbers[row * uniforms.dimInner + col];
	return result;
	}
	return 0.;
	}

	fn mm_readB(row : u32, col : u32) -> f32 {
	if (row < uniforms.dimInner && col < uniforms.dimBOuter)
	{
	let result : f32 = secondMatrix.numbers[row * uniforms.dimBOuter + col];
	return result;
	}
	return 0.;
	}

	fn mm_write(row : u32, col : u32, value : f32) {
	if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)
	{
	let index : u32 = col + row * uniforms.dimBOuter;
	resultMatrix.numbers[index] = value;
	}
	}

	const RowPerThread : u32 = 4u;
	const ColPerThread : u32 = 4u;
	const TileAOuter : u32 = 32u;
	const TileBOuter : u32 = 32u;
	const TileInner : u32 = 32u;)";

	const std::string& kMatMulFloatSharedArray1D = R"(
	var<workgroup> mm_Asub : array<f32, 1024>;
	var<workgroup> mm_Bsub : array<f32, 1024>;)";
	const std::string& kMatMulFloatSharedArray2D = R"(
	var<workgroup> mm_Asub : array<array<f32, 32>, 32>;
	var<workgroup> mm_Bsub : array<array<f32, 32>, 32>;)";
	const std::string& kMatMulFloatBodyPart1 = R"(
	@compute @workgroup_size(8, 8, 1)
	fn main(@builtin(local_invocation_id) local_id : vec3u,
	@builtin(global_invocation_id) global_id : vec3u) {
	let tileRow : u32 = local_id.y * RowPerThread;
	let tileCol : u32 = local_id.x * ColPerThread;

	let globalRow : u32 = global_id.y * RowPerThread;
	let globalCol : u32 = global_id.x * ColPerThread;

	let numTiles : u32 = (uniforms.dimInner - 1u) / TileInner + 1u;

	var acc: array<f32, 16>;
	var ACached : f32;
	var BCached : array<f32, 4>;

	let ColPerThreadA : u32 = TileInner / 8u;
	let tileColA : u32 = local_id.x * ColPerThreadA;
	let RowPerThreadB : u32 = TileInner / 8u;
	let tileRowB : u32 = local_id.y * RowPerThreadB;

	// Loop over shared dimension.
	for (var t : u32 = 0u; t < numTiles; t = t + 1u) {
	// Load one tile of A into local memory.
	for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
	for (var innerCol : u32 = 0u; innerCol < ColPerThreadA; innerCol = innerCol + 1u) {
	let inputRow : u32 = tileRow + innerRow;
	let inputCol : u32 = tileColA + innerCol;)";
	const std::string& kMatMulFloatBodyPart2Array1D = R"(
	let index : u32 = inputRow * TileInner + inputCol;
	mm_Asub[index] = mm_readA(globalRow + innerRow, t * TileInner + inputCol);
	}
	}
	// Load one tile of B into local memory.
	for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
	for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
	let inputRow : u32 = tileRowB + innerRow;
	let inputCol : u32 = tileCol + innerCol;
	let index : u32 = inputRow * TileBOuter + inputCol;

	mm_Bsub[index] = mm_readB(t * TileInner + inputRow, globalCol + innerCol);;
	}
	}

	workgroupBarrier();

	// Compute acc values for a single thread.
	for (var k : u32 = 0u; k < TileInner; k = k + 1u) {
	for (var inner : u32 = 0u; inner < ColPerThread; inner = inner + 1u) {
	BCached[inner] = mm_Bsub[k * TileBOuter + tileCol + inner];
	}

	for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
	ACached = mm_Asub[(tileRow + innerRow) * TileInner + k];)";
	const std::string& kMatMulFloatBodyPart2Array2D = R"(
	mm_Asub[inputRow][inputCol] = mm_readA(globalRow + innerRow, t * TileInner + inputCol);
	}
	}
	// Load one tile of B into local memory.
	for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
	for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
	let inputRow : u32 = tileRowB + innerRow;
	let inputCol : u32 = tileCol + innerCol;

	mm_Bsub[innerCol][inputCol] = mm_readB(t * TileInner + inputRow, globalCol + innerCol);;
	}
	}

	workgroupBarrier();

	// Compute acc values for a single thread.
	for (var k : u32 = 0u; k < TileInner; k = k + 1u) {
	for (var inner : u32 = 0u; inner < ColPerThread; inner = inner + 1u) {
	BCached[inner] = mm_Bsub[k][tileCol + inner];
	}

	for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
	ACached = mm_Asub[tileRow + innerRow][k];)";
	const std::string& kMatMulFloatBodyPart3 = R"(
	for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
	let index : u32 = innerRow * ColPerThread + innerCol;
	acc[index] = acc[index] + ACached * BCached[innerCol];
	}
	}
	}

	workgroupBarrier();
	}

	for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
	for (var innerCol : u32 = 0u; innerCol < ColPerThread; innerCol = innerCol + 1u) {
	let index : u32 = innerRow * ColPerThread + innerCol;
	mm_write(globalRow + innerRow,
	globalCol + innerCol,
	acc[index]);
	}
	}
	})";
	const std::string& kMatMulFloatOneDimensionalSharedArray =
	kMatMulFloatHeader + kMatMulFloatSharedArray1D + kMatMulFloatBodyPart1 +
	kMatMulFloatBodyPart2Array1D + kMatMulFloatBodyPart3;

	const std::string& kMatMulFloatTwoDimensionalSharedArray =
	kMatMulFloatHeader + kMatMulFloatSharedArray2D + kMatMulFloatBodyPart1 +
	kMatMulFloatBodyPart2Array2D + kMatMulFloatBodyPart3;

	// The vec4 version requires that dimInner and dimBOuter are divisible by 4.
	const std::string& kMatMulVec4Header = R"(
	struct Uniforms {
	dimAOuter : u32,
	dimInner : u32,
	dimBOuter : u32,
	}
	struct Matrix {
	numbers: array<vec4f>
	}

	@group(0) @binding(0) var<storage, read> firstMatrix : Matrix;
	@group(0) @binding(1) var<storage, read> secondMatrix : Matrix;
	@group(0) @binding(2) var<storage, read_write> resultMatrix : Matrix;
	@group(0) @binding(3) var<uniform> uniforms : Uniforms;

	fn mm_readA(row : u32, col : u32) -> vec4f {
	if (row < uniforms.dimAOuter && col < uniforms.dimInner)
	{
	let result : vec4f = firstMatrix.numbers[row * uniforms.dimInner / 4u + col];
	return result;
	}
	return vec4f(0.0, 0.0, 0.0, 0.0);
	}

	fn mm_readB(row : u32, col : u32) -> vec4f {
	if (row < uniforms.dimInner && col < uniforms.dimBOuter)
	{
	let result : vec4f = secondMatrix.numbers[row * uniforms.dimBOuter / 4u + col];
	return result;
	}
	return vec4f(0.0, 0.0, 0.0, 0.0);
	}

	fn mm_write(row : u32, col : u32, value : vec4f) {
	if (row < uniforms.dimAOuter && col < uniforms.dimBOuter)
	{
	let index : u32 = col + row * uniforms.dimBOuter / 4u;
	resultMatrix.numbers[index] = value;
	}
	}

	const RowPerThread : u32 = 4u;
	const ColPerThread : u32 = 4u;
	const TileOuter : u32 = 32u;
	const TileInner : u32 = 32u;)";
	const std::string& kMatMulVec4SharedArray1D = R"(
	var<workgroup> mm_Asub : array<vec4f, 256>;
	var<workgroup> mm_Bsub : array<vec4f, 256>;)";
	const std::string& kMatMulVec4SharedArray2D = R"(
	var<workgroup> mm_Asub : array<array<vec4f, 8>, 32>;
	var<workgroup> mm_Bsub : array<array<vec4f, 8>, 32>;)";
	const std::string& kMatMulVec4BodyPart1 = R"(
	@compute @workgroup_size(8, 8, 1)
	fn main(@builtin(local_invocation_id) local_id : vec3u,
	@builtin(global_invocation_id) global_id : vec3u) {
	let tileRow : u32 = local_id.y * RowPerThread;
	let tileCol : u32 = local_id.x;

	let globalRow : u32 = global_id.y * RowPerThread;
	let globalCol : u32 = global_id.x;

	let numTiles : u32 = (uniforms.dimInner - 1u) / TileInner + 1u;

	var acc: array<vec4f, 4>;
	var ACached : vec4f;
	var BCached : array<vec4f, 4>;

	var globalColA : u32 = tileCol;
	let RowPerThreadB : u32 = TileInner / 8u;
	let tileRowB : u32 = local_id.y * RowPerThreadB;

	// Loop over shared dimension.
	for (var t : u32 = 0u; t < numTiles; t = t + 1u) {
	// Load one tile of A into local memory.
	for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
	let inputRow : u32 = tileRow + innerRow;
	let inputCol : u32 = tileCol;)";
	const std::string& kMatMulVec4BodyPart2Array1D = R"(
	let index : u32 = inputRow * TileInner / ColPerThread + inputCol;
	mm_Asub[index] = mm_readA(globalRow + innerRow, globalColA);
	}
	globalColA = globalColA + TileInner / ColPerThread;

	// Load one tile of B into local memory.
	for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
	let inputRow : u32 = tileRowB + innerRow;
	let inputCol : u32 = tileCol;
	let index : u32 = inputRow * TileOuter / ColPerThread + inputCol;
	mm_Bsub[index] = mm_readB(t * TileInner + inputRow, globalCol);;
	}

	workgroupBarrier();

	// Compute acc values for a single thread.
	for (var k : u32 = 0u; k < TileInner / ColPerThread; k = k + 1u) {
	BCached[0] = mm_Bsub[(k * ColPerThread) * (TileOuter / ColPerThread) + tileCol];
	BCached[1] = mm_Bsub[(k * ColPerThread + 1u) * (TileOuter / ColPerThread) + tileCol];
	BCached[2] = mm_Bsub[(k * ColPerThread + 2u) * (TileOuter / ColPerThread) + tileCol];
	BCached[3] = mm_Bsub[(k * ColPerThread + 3u) * (TileOuter / ColPerThread) + tileCol];

	for (var i : u32 = 0u; i < RowPerThread; i = i + 1u) {
	ACached = mm_Asub[(tileRow + i) * (TileInner / ColPerThread) + k];)";
	const std::string& kMatMulVec4BodyPart2Array2D = R"(
	mm_Asub[inputRow][inputCol] = mm_readA(globalRow + innerRow, globalColA);
	}
	globalColA = globalColA + TileInner / ColPerThread;

	// Load one tile of B into local memory.
	for (var innerRow : u32 = 0u; innerRow < RowPerThreadB; innerRow = innerRow + 1u) {
	let inputRow : u32 = tileRowB + innerRow;
	let inputCol : u32 = tileCol;
	mm_Bsub[inputRow][inputCol] = mm_readB(t * TileInner + inputRow, globalCol);;
	}

	workgroupBarrier();

	// Compute acc values for a single thread.
	for (var k : u32 = 0u; k < TileInner / ColPerThread; k = k + 1u) {
	BCached[0] = mm_Bsub[k * ColPerThread][tileCol];
	BCached[1] = mm_Bsub[k * ColPerThread + 1u][tileCol];
	BCached[2] = mm_Bsub[k * ColPerThread + 2u][tileCol];
	BCached[3] = mm_Bsub[k * ColPerThread + 3u][tileCol];

	for (var i : u32 = 0u; i < RowPerThread; i = i + 1u) {
	ACached = mm_Asub[tileRow + i][k];)";
	const std::string& kMatMulVec4BodyPart3 = R"(
	acc[i] = BCached[0] * ACached.x + acc[i];
	acc[i] = BCached[1] * ACached.y + acc[i];
	acc[i] = BCached[2] * ACached.z + acc[i];
	acc[i] = BCached[3] * ACached.w + acc[i];
	}
	}

	workgroupBarrier();
	}

	for (var innerRow : u32 = 0u; innerRow < RowPerThread; innerRow = innerRow + 1u) {
	mm_write(globalRow + innerRow,
	globalCol,
	acc[innerRow]);
	}
	})";

	const std::string& kMatMulVec4OneDimensionalSharedArray =
	kMatMulVec4Header + kMatMulVec4SharedArray1D + kMatMulVec4BodyPart1 +
	kMatMulVec4BodyPart2Array1D + kMatMulVec4BodyPart3;

	const std::string& kMatMulVec4TwoDimensionalSharedArray =
	kMatMulVec4Header + kMatMulVec4SharedArray2D + kMatMulVec4BodyPart1 +
	kMatMulVec4BodyPart2Array2D + kMatMulVec4BodyPart3;

	constexpr unsigned int kNumIterations = 50;

	enum class MatMulMethod {
	MatMulFloatOneDimSharedArray,
	MatMulFloatTwoDimSharedArray,
	MatMulVec4OneDimSharedArray,
	MatMulVec4TwoDimSharedArray
	};

	std::ostream& operator<<(std::ostream& ostream, const MatMulMethod& matMulMethod) {
	switch (matMulMethod) {
	case MatMulMethod::MatMulFloatOneDimSharedArray:
	ostream << "MatMulFloatOneDimSharedArray";
	break;
	case MatMulMethod::MatMulFloatTwoDimSharedArray:
	ostream << "MatMulFloatTwoDimSharedArray";
	break;
	case MatMulMethod::MatMulVec4OneDimSharedArray:
	ostream << "MatMulVec4OneDimSharedArray";
	break;
	case MatMulMethod::MatMulVec4TwoDimSharedArray:
	ostream << "MatMulVec4TwoDimSharedArray";
	break;
	}
	return ostream;
	}

	using DimAOuter = uint32_t;
	using DimInner = uint32_t;
	using DimBOuter = uint32_t;
	DAWN_TEST_PARAM_STRUCT(ShaderRobustnessParams, MatMulMethod, DimAOuter, DimInner, DimBOuter);

	// Test the execution time of matrix multiplication (A [dimAOuter, dimInner] * B [dimInner,
	// dimBOuter]) on the GPU and see the difference between robustness on and off.
	class ShaderRobustnessPerf : public DawnPerfTestWithParams<ShaderRobustnessParams> {
	public:
	ShaderRobustnessPerf()
	: DawnPerfTestWithParams(kNumIterations, 1),
	mDimAOuter(GetParam().mDimAOuter),
	mDimInner(GetParam().mDimInner),
	mDimBOuter(GetParam().mDimBOuter) {}
	~ShaderRobustnessPerf() override = default;

	void SetUp() override;

	private:
	void Step() override;

	wgpu::BindGroup mBindGroup;
	wgpu::ComputePipeline mPipeline;
	uint32_t mDimAOuter;
	uint32_t mDimInner;
	uint32_t mDimBOuter;
	};

	void ShaderRobustnessPerf::SetUp() {
	DawnPerfTestWithParams<ShaderRobustnessParams>::SetUp();

	const size_t dataASize = mDimAOuter * mDimInner;
	std::vector<float> dataA(dataASize);
	uint64_t byteASize = sizeof(float) * dataA.size();
	// It's ok to use all zeros to do the matrix multiplication for performance test.
	wgpu::Buffer bufA =
	utils::CreateBufferFromData(device, dataA.data(), byteASize, wgpu::BufferUsage::Storage);

	const size_t dataBSize = mDimInner * mDimBOuter;
	std::vector<float> dataB(dataBSize);
	uint64_t byteBSize = sizeof(float) * dataB.size();
	wgpu::Buffer bufB =
	utils::CreateBufferFromData(device, dataB.data(), byteBSize, wgpu::BufferUsage::Storage);

	uint64_t byteDstSize = sizeof(float) * mDimAOuter * mDimBOuter;
	wgpu::BufferDescriptor desc = {};
	desc.usage = wgpu::BufferUsage::Storage;
	desc.size = byteDstSize;
	wgpu::Buffer dst = device.CreateBuffer(&desc);

	uint32_t uniformData[] = {mDimAOuter, mDimInner, mDimBOuter};
	wgpu::Buffer uniformBuffer = utils::CreateBufferFromData(
	device, uniformData, sizeof(uniformData), wgpu::BufferUsage::Uniform);

	wgpu::ShaderModule module;
	switch (GetParam().mMatMulMethod) {
	case MatMulMethod::MatMulFloatOneDimSharedArray: {
	module =
	utils::CreateShaderModule(device, kMatMulFloatOneDimensionalSharedArray.c_str());
	break;
	}

	case MatMulMethod::MatMulFloatTwoDimSharedArray: {
	module =
	utils::CreateShaderModule(device, kMatMulFloatTwoDimensionalSharedArray.c_str());
	break;
	}

	case MatMulMethod::MatMulVec4OneDimSharedArray: {
	module =
	utils::CreateShaderModule(device, kMatMulVec4OneDimensionalSharedArray.c_str());
	break;
	}

	case MatMulMethod::MatMulVec4TwoDimSharedArray: {
	module =
	utils::CreateShaderModule(device, kMatMulVec4TwoDimensionalSharedArray.c_str());
	break;
	}
	}

	wgpu::ComputePipelineDescriptor csDesc;
	csDesc.compute.module = module;
	csDesc.compute.entryPoint = "main";
	mPipeline = device.CreateComputePipeline(&csDesc);

	mBindGroup = utils::MakeBindGroup(device, mPipeline.GetBindGroupLayout(0),
	{
	{0, bufA, 0, byteASize},
	{1, bufB, 0, byteBSize},
	{2, dst, 0, byteDstSize},
	{3, uniformBuffer, 0, sizeof(uniformData)},
	});
	}

	void ShaderRobustnessPerf::Step() {
	wgpu::CommandBuffer commands;
	{
	wgpu::CommandEncoder encoder = device.CreateCommandEncoder();
	if (SupportsTimestampQuery()) {
	RecordBeginTimestamp(encoder);
	}
	wgpu::ComputePassEncoder pass = encoder.BeginComputePass();
	pass.SetPipeline(mPipeline);
	pass.SetBindGroup(0, mBindGroup);
	for (unsigned int i = 0; i < kNumIterations; ++i) {
	pass.DispatchWorkgroups(ceil(static_cast<float>(mDimBOuter) / float{kTileSize}),
	ceil(static_cast<float>(mDimAOuter) / float{kTileSize}), 1);
	}
	pass.End();
	if (SupportsTimestampQuery()) {
	RecordEndTimestampAndResolveQuerySet(encoder);
	}

	commands = encoder.Finish();
	}

	queue.Submit(1, &commands);

	if (SupportsTimestampQuery()) {
	ComputeGPUElapsedTime();
	}
	}

	TEST_P(ShaderRobustnessPerf, Run) {
	RunTest();
	}

	DAWN_INSTANTIATE_TEST_P(ShaderRobustnessPerf,
	{D3D12Backend(), D3D12Backend({"disable_robustness"}, {}), MetalBackend(),
	MetalBackend({"disable_robustness"}, {}), OpenGLBackend(),
	OpenGLBackend({"disable_robustness"}, {}), VulkanBackend(),
	VulkanBackend({"disable_robustness"}, {})},
	{MatMulMethod::MatMulFloatOneDimSharedArray,
	MatMulMethod::MatMulFloatTwoDimSharedArray,
	MatMulMethod::MatMulVec4OneDimSharedArray,
	MatMulMethod::MatMulVec4TwoDimSharedArray},
	{512u},
	{512u},
	{512u});

	} // anonymous namespace
	} // namespace dawn