David Monahan | bd73808 | 2023-12-08 12:50:02 +0000 | [diff] [blame^] | 1 | // |
| 2 | // Copyright © 2024 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "GpuFsaConvolution2d.hpp" |
| 7 | |
| 8 | #include <armnn/Types.hpp> |
| 9 | |
| 10 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 11 | |
| 12 | #include <arm_compute/core/ITensorInfo.h> |
| 13 | #include <arm_compute/core/TensorInfo.h> |
| 14 | #include <arm_compute/core/TensorShape.h> |
| 15 | #include <arm_compute/core/CL/CLKernelLibrary.h> |
| 16 | #include <arm_compute/core/CL/CLCompileContext.h> |
| 17 | |
| 18 | #include <arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h> |
| 19 | #include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h> |
| 20 | #include <src/dynamic_fusion/sketch/gpu/GpuWorkloadContextImpl.h> |
| 21 | #include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h> |
| 22 | #include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h> |
| 23 | |
| 24 | #include <vector> |
| 25 | #include <iostream> |
| 26 | |
| 27 | namespace armnn |
| 28 | { |
| 29 | |
| 30 | using namespace armcomputetensorutils; |
| 31 | |
| 32 | arm_compute::Status GpuFsaConvolution2dValidate(const TensorInfo& input, |
| 33 | const Convolution2dDescriptor& descriptor, |
| 34 | const TensorInfo& weights, |
| 35 | const Optional<TensorInfo>& biases) |
| 36 | { |
| 37 | // Create a new workload sketch, for validation purposes |
| 38 | auto compileCtx = arm_compute::CLKernelLibrary::get().get_compile_context(); |
| 39 | auto workloadContext = GpuWorkloadContext(&compileCtx); |
| 40 | GpuWorkloadSketch sketch{ &workloadContext }; |
| 41 | |
| 42 | // Build and create tensor infos using the sketch |
| 43 | const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); |
| 44 | arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); |
| 45 | aclWeightsInfo.set_are_values_constant(weights.IsConstant()); |
| 46 | |
| 47 | auto inputInfo = workloadContext.create_tensor_info(aclInputInfo); |
| 48 | auto weightInfo = workloadContext.create_tensor_info(aclWeightsInfo); |
| 49 | |
| 50 | // Only create the bias tensor info if enabled, otherwise pass nullptr to validate_op |
| 51 | arm_compute::TensorInfo aclBiasInfo; |
| 52 | arm_compute::TensorInfo biasSketchInfo; |
| 53 | arm_compute::TensorInfo* biasSketchInfoPtr = nullptr; |
| 54 | |
| 55 | if (descriptor.m_BiasEnabled) |
| 56 | { |
| 57 | if(!biases.has_value()) |
| 58 | { |
| 59 | throw InvalidArgumentException("GpuFsaConvolution2d::ValidateOp: No biases set when biases are enabled"); |
| 60 | } |
| 61 | aclBiasInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); |
| 62 | aclBiasInfo.set_are_values_constant(biases.value().IsConstant()); |
| 63 | |
| 64 | biasSketchInfo = workloadContext.create_tensor_info(aclBiasInfo); |
| 65 | biasSketchInfoPtr = &biasSketchInfo; |
| 66 | } |
| 67 | |
| 68 | // Set Conv2d attributes using descriptor |
| 69 | const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX, |
| 70 | descriptor.m_DilationY); |
| 71 | const arm_compute::Padding2D aclPadInfo = BuildArmComputePaddingInfo(descriptor); |
| 72 | const arm_compute::Size2D aclStrideInfo = BuildArmComputeSize2D(descriptor.m_StrideX, descriptor.m_StrideY); |
| 73 | |
| 74 | Conv2dAttributes conv2DAttributes{}; |
| 75 | conv2DAttributes.dilation(aclDilationInfo); |
| 76 | conv2DAttributes.pad(aclPadInfo); |
| 77 | conv2DAttributes.stride(aclStrideInfo); |
| 78 | |
| 79 | // Validate operator, check status and update reasonIfUnsupported |
| 80 | arm_compute::Status aclStatus = GpuConv2d::validate_op(sketch, |
| 81 | &inputInfo, |
| 82 | &weightInfo, |
| 83 | biasSketchInfoPtr, |
| 84 | conv2DAttributes); |
| 85 | |
| 86 | return aclStatus; |
| 87 | } |
| 88 | |
| 89 | void GpuFsaConvolution2dCreateOp(GpuFsaPreCompiledBlob* blob, |
| 90 | const TensorInfo& input, |
| 91 | const Convolution2dDescriptor& descriptor, |
| 92 | const TensorInfo& weights, |
| 93 | const Optional<TensorInfo>& biases) |
| 94 | { |
| 95 | /* |
| 96 | * Creating an Op for the GpuFds backend requires us to create and maintain quite a bit of data, which is then stored |
| 97 | * in a GpuFsaPreCompiledBlob for execution later. Specifically we need: |
| 98 | * GpuWorkloadContext, this contains the TensorInfos and is unique to the Graph being executed |
| 99 | * Sketch, this is similar to a subgraph and can contain one or more operations. Multiple ops can be "fused" together |
| 100 | * using a single sketch. |
| 101 | * The TensorInfoIds, these are the ids of the TensorInfos used when creating the sketch. They refer to the TensorInfos |
| 102 | * stored within the GpuWorkloadContext and are used to fetch them later when executing the sketch. |
| 103 | */ |
| 104 | using namespace arm_compute::experimental::dynamic_fusion; |
| 105 | GpuWorkloadSketch* sketch = blob->sketch.get(); |
| 106 | GpuWorkloadContext* workloadContext = blob->workloadContext.get(); |
| 107 | std::vector<int32_t> inputIds = {}; |
| 108 | std::vector<int32_t> outputIds = {}; |
| 109 | |
| 110 | // Build and create tensor infos using the sketch |
| 111 | const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); |
| 112 | arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); |
| 113 | aclWeightsInfo.set_are_values_constant(weights.IsConstant()); |
| 114 | auto inputInfo = workloadContext->create_tensor_info(aclInputInfo); |
| 115 | aclWeightsInfo.set_are_values_constant(weights.IsConstant()); |
| 116 | inputIds.emplace_back(inputInfo.id()); |
| 117 | |
| 118 | auto weightInfo = workloadContext->create_tensor_info(aclWeightsInfo); |
| 119 | inputIds.emplace_back(weightInfo.id()); |
| 120 | |
| 121 | // Only create the bias tensor info if enabled, otherwise pass nullptr to validate_op |
| 122 | arm_compute::TensorInfo aclBiasInfo; |
| 123 | arm_compute::TensorInfo biasSketchInfo; |
| 124 | arm_compute::ITensorInfo* biasSketchInfoPtr = nullptr; |
| 125 | |
| 126 | if (descriptor.m_BiasEnabled) |
| 127 | { |
| 128 | if(!biases.has_value()) |
| 129 | { |
| 130 | throw InvalidArgumentException("GpuFsaConvolution2d::CreateOp: No biases set when biases are enabled"); |
| 131 | } |
| 132 | aclBiasInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); |
| 133 | aclBiasInfo.set_are_values_constant(biases.value().IsConstant()); |
| 134 | |
| 135 | biasSketchInfo = workloadContext->create_tensor_info(aclBiasInfo); |
| 136 | inputIds.emplace_back(biasSketchInfo.id()); |
| 137 | biasSketchInfoPtr = workloadContext->implementation().get_tensor_info(biasSketchInfo.id()); |
| 138 | } |
| 139 | |
| 140 | // Set Conv2d attributes using descriptor |
| 141 | const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX, |
| 142 | descriptor.m_DilationY); |
| 143 | const arm_compute::Padding2D aclPadInfo = BuildArmComputePaddingInfo(descriptor); |
| 144 | const arm_compute::Size2D aclStrideInfo = BuildArmComputeSize2D(descriptor.m_StrideX, descriptor.m_StrideY); |
| 145 | |
| 146 | Conv2dAttributes conv2DAttributes{}; |
| 147 | conv2DAttributes.dilation(aclDilationInfo); |
| 148 | conv2DAttributes.pad(aclPadInfo); |
| 149 | conv2DAttributes.stride(aclStrideInfo); |
| 150 | |
| 151 | // Validate operator, check status and update reasonIfUnsupported |
| 152 | arm_compute::Status aclStatus = |
| 153 | GpuConv2d::validate_op(*sketch, |
| 154 | workloadContext->implementation().get_tensor_info(inputInfo.id()), |
| 155 | workloadContext->implementation().get_tensor_info(weightInfo.id()), |
| 156 | biasSketchInfoPtr, |
| 157 | conv2DAttributes); |
| 158 | |
| 159 | const bool supported = (aclStatus.error_code() == arm_compute::ErrorCode::OK); |
| 160 | if (!supported) |
| 161 | { |
| 162 | throw BackendCapabilityException("\"GpuFsa\" backend failed during Convolution2D operation validation"); |
| 163 | } |
| 164 | |
| 165 | arm_compute::ITensorInfo* convOutInfo = |
| 166 | GpuConv2d::create_op(*sketch, |
| 167 | workloadContext->implementation().get_tensor_info(inputInfo.id()), |
| 168 | workloadContext->implementation().get_tensor_info(weightInfo.id()), |
| 169 | biasSketchInfoPtr, |
| 170 | conv2DAttributes); |
| 171 | |
| 172 | arm_compute::TensorInfo outputDstInfo = workloadContext->create_tensor_info(); |
| 173 | outputIds.emplace_back(outputDstInfo.id()); |
| 174 | |
| 175 | GpuOutput::create_op(*sketch, convOutInfo, workloadContext->implementation().get_tensor_info(outputDstInfo.id())); |
| 176 | blob->inputIds = std::make_unique<std::vector<int32_t>>(inputIds); |
| 177 | blob->outputIds = std::make_unique<std::vector<int32_t>>(outputIds); |
| 178 | } |
| 179 | |
| 180 | } // namespace armnn |