diff options
Diffstat (limited to 'src/backends/gpuFsa/layerValidators/GpuFsaConvolution2dValidate.cpp')
| -rw-r--r-- | src/backends/gpuFsa/layerValidators/GpuFsaConvolution2dValidate.cpp | 90 |
1 files changed, 64 insertions, 26 deletions
diff --git a/src/backends/gpuFsa/layerValidators/GpuFsaConvolution2dValidate.cpp b/src/backends/gpuFsa/layerValidators/GpuFsaConvolution2dValidate.cpp index 2b6c2ee3dc..269442b60c 100644 --- a/src/backends/gpuFsa/layerValidators/GpuFsaConvolution2dValidate.cpp +++ b/src/backends/gpuFsa/layerValidators/GpuFsaConvolution2dValidate.cpp @@ -19,8 +19,8 @@ #include <arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h> #include <arm_compute/dynamic_fusion/sketch/OperatorAttributes.h> #include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h> -#include <arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h> #include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h> +#include <arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h> #include <vector> #include <iostream> @@ -30,29 +30,24 @@ namespace armnn using namespace armcomputetensorutils; -arm_compute::Status GpuFsaConvolution2dValidate(const TensorInfo& input, - const Convolution2dDescriptor& descriptor, - const TensorInfo& weights, - const Optional<TensorInfo>& biases) +inline arm_compute::Status ValidateAndCreateOp(const TensorInfo& input, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const Optional<TensorInfo>& biases, + GpuWorkloadSketch& sketch, + const bool createOp = false) { - using namespace arm_compute::experimental::dynamic_fusion; - - // Create a new workload sketch, for validation purposes - auto compileCtx = arm_compute::CLKernelLibrary::get().get_compile_context(); - auto gpuCtx = GpuWorkloadContext(&compileCtx); - GpuWorkloadSketch sketch{ &gpuCtx }; - // Build and create tensor infos using the sketch - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); + const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); + arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); aclWeightsInfo.set_are_values_constant(weights.IsConstant()); auto inputInfo = sketch.create_tensor_info(aclInputInfo); auto weightInfo = sketch.create_tensor_info(aclWeightsInfo); // Only create the bias tensor info if enabled, otherwise pass nullptr to validate_op - arm_compute::TensorInfo aclBiasInfo; - arm_compute::TensorInfo biasSketchInfo; + arm_compute::TensorInfo aclBiasInfo; + arm_compute::TensorInfo biasSketchInfo; arm_compute::TensorInfo* biasSketchInfoPtr = nullptr; if (descriptor.m_BiasEnabled) @@ -61,28 +56,71 @@ arm_compute::Status GpuFsaConvolution2dValidate(const TensorInfo& input, aclBiasInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); aclBiasInfo.set_are_values_constant(biases.value().IsConstant()); - biasSketchInfo = sketch.create_tensor_info(aclBiasInfo); + biasSketchInfo = sketch.create_tensor_info(aclBiasInfo); biasSketchInfoPtr = &biasSketchInfo; } // Set Conv2d attributes using descriptor - const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX, descriptor.m_DilationY); - const arm_compute::Padding2D aclPadInfo = BuildArmComputePaddingInfo(descriptor); - const arm_compute::Size2D aclStrideInfo = BuildArmComputeSize2D(descriptor.m_StrideX, descriptor.m_StrideY); + const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX, + descriptor.m_DilationY); + const arm_compute::Padding2D aclPadInfo = BuildArmComputePaddingInfo(descriptor); + const arm_compute::Size2D aclStrideInfo = BuildArmComputeSize2D(descriptor.m_StrideX, descriptor.m_StrideY); Conv2dAttributes conv2DAttributes{}; conv2DAttributes.dilation(aclDilationInfo); conv2DAttributes.pad(aclPadInfo); conv2DAttributes.stride(aclStrideInfo); + // Validate operator, check status and update reasonIfUnsupported + arm_compute::Status aclStatus = GpuConv2d::validate_op(sketch, + &inputInfo, + &weightInfo, + biasSketchInfoPtr, + conv2DAttributes); + + if (createOp) { - // Validate operator, check status and update reasonIfUnsupported - return GpuConv2d::validate_op(sketch, - &inputInfo, - &weightInfo, - biasSketchInfoPtr, - conv2DAttributes); + const bool supported = (aclStatus.error_code() == arm_compute::ErrorCode::OK); + if (!supported) + { + throw BackendCapabilityException("\"GpuFsa\" backend failed during operation validation when attempting " + "to fuse a GpuConv2d operator into the existing workload sketch."); + } + + arm_compute::ITensorInfo* convOutInfo = GpuConv2d::create_op(sketch, + &inputInfo, + &weightInfo, + biasSketchInfoPtr, + conv2DAttributes); + + // Temporary fix until fusing attempt is make for GpuFsa backend and Output layer workload is created. + auto outputInfo = sketch.create_tensor_info(); + GpuOutput::create_op(sketch, convOutInfo, &outputInfo); } + + return aclStatus; +} + +arm_compute::Status GpuFsaConvolution2dValidate(const TensorInfo& input, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const Optional<TensorInfo>& biases) +{ + // Create a new workload sketch, for validation purposes + auto compileCtx = arm_compute::CLKernelLibrary::get().get_compile_context(); + auto gpuCtx = GpuWorkloadContext(&compileCtx); + GpuWorkloadSketch sketch{ &gpuCtx }; + + return ValidateAndCreateOp(input, descriptor, weights, biases, sketch); +} + +void GpuFsaConvolution2dCreateOp(const TensorInfo& input, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const Optional<TensorInfo>& biases, + GpuWorkloadSketch& sketch) +{ + ValidateAndCreateOp(input, descriptor, weights, biases, sketch, true); } } // namespace armnn
\ No newline at end of file |