diff options
Diffstat (limited to 'src/backends/cl/workloads')
95 files changed, 0 insertions, 5314 deletions
diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt deleted file mode 100644 index de62ca9496..0000000000 --- a/src/backends/cl/workloads/CMakeLists.txt +++ /dev/null @@ -1,107 +0,0 @@ -# -# Copyright © 2017 Arm Ltd. All rights reserved. -# SPDX-License-Identifier: MIT -# - -list(APPEND armnnClBackendWorkloads_sources - ClAbsWorkload.cpp - ClAbsWorkload.hpp - ClActivationWorkload.cpp - ClActivationWorkload.hpp - ClAdditionWorkload.cpp - ClAdditionWorkload.hpp - ClArgMinMaxWorkload.cpp - ClArgMinMaxWorkload.hpp - ClBatchNormalizationFloatWorkload.cpp - ClBatchNormalizationFloatWorkload.hpp - ClBatchToSpaceNdWorkload.cpp - ClBatchToSpaceNdWorkload.hpp - ClConcatWorkload.cpp - ClConcatWorkload.hpp - ClConstantWorkload.cpp - ClConstantWorkload.hpp - ClConvertFp16ToFp32Workload.cpp - ClConvertFp16ToFp32Workload.hpp - ClConvertFp32ToFp16Workload.cpp - ClConvertFp32ToFp16Workload.hpp - ClConvolution2dWorkload.cpp - ClConvolution2dWorkload.hpp - ClDepthToSpaceWorkload.cpp - ClDepthToSpaceWorkload.hpp - ClDepthwiseConvolutionWorkload.cpp - ClDepthwiseConvolutionWorkload.hpp - ClDequantizeWorkload.cpp - ClDequantizeWorkload.hpp - ClDivisionFloatWorkload.cpp - ClDivisionFloatWorkload.hpp - ClFloorFloatWorkload.cpp - ClFloorFloatWorkload.hpp - ClFullyConnectedWorkload.cpp - ClFullyConnectedWorkload.hpp - ClGreaterWorkload.cpp - ClGreaterWorkload.hpp - ClInstanceNormalizationWorkload.cpp - ClInstanceNormalizationWorkload.hpp - ClL2NormalizationFloatWorkload.cpp - ClL2NormalizationFloatWorkload.hpp - ClLstmFloatWorkload.cpp - ClLstmFloatWorkload.hpp - ClMaximumWorkload.cpp - ClMaximumWorkload.hpp - ClMeanWorkload.cpp - ClMeanWorkload.hpp - ClMinimumWorkload.cpp - ClMinimumWorkload.hpp - ClMultiplicationWorkload.cpp - ClMultiplicationWorkload.hpp - ClNormalizationFloatWorkload.cpp - ClNormalizationFloatWorkload.hpp - ClPadWorkload.cpp - ClPadWorkload.hpp - ClPermuteWorkload.cpp - ClPermuteWorkload.hpp - ClPooling2dWorkload.cpp - ClPooling2dWorkload.hpp - ClPreluWorkload.cpp - ClPreluWorkload.hpp - ClQuantizedLstmWorkload.cpp - ClQuantizedLstmWorkload.hpp - ClQuantizeWorkload.cpp - ClQuantizeWorkload.hpp - ClReshapeWorkload.cpp - ClReshapeWorkload.hpp - ClResizeWorkload.cpp - ClResizeWorkload.hpp - ClRsqrtWorkload.cpp - ClRsqrtWorkload.hpp - ClSliceWorkload.cpp - ClSliceWorkload.hpp - ClSoftmaxBaseWorkload.cpp - ClSoftmaxBaseWorkload.hpp - ClSoftmaxFloatWorkload.cpp - ClSoftmaxFloatWorkload.hpp - ClSoftmaxUint8Workload.cpp - ClSoftmaxUint8Workload.hpp - ClSpaceToBatchNdWorkload.hpp - ClSpaceToBatchNdWorkload.cpp - ClSpaceToDepthWorkload.cpp - ClSpaceToDepthWorkload.hpp - ClSplitterWorkload.cpp - ClSplitterWorkload.hpp - ClStackWorkload.cpp - ClStackWorkload.hpp - ClStridedSliceWorkload.cpp - ClStridedSliceWorkload.hpp - ClSubtractionWorkload.cpp - ClSubtractionWorkload.hpp - ClTransposeConvolution2dWorkload.cpp - ClTransposeConvolution2dWorkload.hpp - ClWorkloads.hpp - ClWorkloadUtils.hpp -) - -add_library(armnnClBackendWorkloads OBJECT ${armnnClBackendWorkloads_sources}) -target_include_directories(armnnClBackendWorkloads PRIVATE ${PROJECT_SOURCE_DIR}/src/armnn) -target_include_directories(armnnClBackendWorkloads PRIVATE ${PROJECT_SOURCE_DIR}/src/armnnUtils) -target_include_directories(armnnClBackendWorkloads PRIVATE ${PROJECT_SOURCE_DIR}/src/backends) -target_include_directories(armnnClBackendWorkloads PRIVATE ${PROJECT_SOURCE_DIR}/src/profiling) diff --git a/src/backends/cl/workloads/ClAbsWorkload.cpp b/src/backends/cl/workloads/ClAbsWorkload.cpp deleted file mode 100644 index 058c453c6b..0000000000 --- a/src/backends/cl/workloads/ClAbsWorkload.cpp +++ /dev/null @@ -1,44 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClAbsWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <cl/ClTensorHandle.hpp> - -#include <boost/cast.hpp> - -namespace armnn -{ - -arm_compute::Status ClAbsWorkloadValidate(const TensorInfo& input, const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - return arm_compute::CLAbsLayer::validate(&aclInput, &aclOutput); -} - -ClAbsWorkload::ClAbsWorkload(const AbsQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<AbsQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClAbsWorkload", 1, 1); - - arm_compute::ICLTensor& input = boost::polymorphic_downcast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = boost::polymorphic_downcast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_AbsLayer.configure(&input, &output); -} - -void ClAbsWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClAbsWorkload_Execute"); - RunClFunction(m_AbsLayer, CHECK_LOCATION()); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClAbsWorkload.hpp b/src/backends/cl/workloads/ClAbsWorkload.hpp deleted file mode 100644 index 763cafcfbd..0000000000 --- a/src/backends/cl/workloads/ClAbsWorkload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/core/Error.h> -#include <arm_compute/runtime/CL/functions/CLElementWiseUnaryLayer.h> - -namespace armnn -{ - -arm_compute::Status ClAbsWorkloadValidate(const TensorInfo& input, const TensorInfo& output); - -class ClAbsWorkload : public BaseWorkload<AbsQueueDescriptor> -{ -public: - ClAbsWorkload(const AbsQueueDescriptor& descriptor, const WorkloadInfo& info); - virtual void Execute() const override; - -private: - mutable arm_compute::CLAbsLayer m_AbsLayer; -}; - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClActivationWorkload.cpp b/src/backends/cl/workloads/ClActivationWorkload.cpp deleted file mode 100644 index 685652036b..0000000000 --- a/src/backends/cl/workloads/ClActivationWorkload.cpp +++ /dev/null @@ -1,52 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClActivationWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <aclCommon/ArmComputeUtils.hpp> - -namespace armnn -{ -arm_compute::Status ClActivationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const ActivationDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - const arm_compute::ActivationLayerInfo activationLayerInfo = - ConvertActivationDescriptorToAclActivationLayerInfo(descriptor); - - return arm_compute::CLActivationLayer::validate(&aclInput, - &aclOutput, - activationLayerInfo); -} - -ClActivationWorkload::ClActivationWorkload(const ActivationQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<ActivationQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClActivationWorkload", 1, 1); - - const arm_compute::ActivationLayerInfo activationLayerInfo = - ConvertActivationDescriptorToAclActivationLayerInfo(m_Data.m_Parameters); - - arm_compute::ICLTensor& input = static_cast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - m_ActivationLayer.configure(&input, &output, activationLayerInfo); -} - -void ClActivationWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClActivationWorkload_Execute"); - RunClFunction(m_ActivationLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClActivationWorkload.hpp b/src/backends/cl/workloads/ClActivationWorkload.hpp deleted file mode 100644 index 35166332e6..0000000000 --- a/src/backends/cl/workloads/ClActivationWorkload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLActivationLayer.h> - -namespace armnn -{ -arm_compute::Status ClActivationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const ActivationDescriptor& descriptor); - -class ClActivationWorkload : public BaseWorkload<ActivationQueueDescriptor> -{ -public: - ClActivationWorkload(const ActivationQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLActivationLayer m_ActivationLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClAdditionWorkload.cpp b/src/backends/cl/workloads/ClAdditionWorkload.cpp deleted file mode 100644 index 18e2400ccd..0000000000 --- a/src/backends/cl/workloads/ClAdditionWorkload.cpp +++ /dev/null @@ -1,54 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClAdditionWorkload.hpp" - -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE; - -ClAdditionWorkload::ClAdditionWorkload(const AdditionQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<AdditionQueueDescriptor>(descriptor, info) -{ - this->m_Data.ValidateInputsOutputs("ClAdditionWorkload", 2, 1); - - arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor(); - m_Layer.configure(&input0, &input1, &output, g_AclConvertPolicy); -} - -void ClAdditionWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClAdditionWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -arm_compute::Status ClAdditionValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0); - const arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - const arm_compute::Status aclStatus = arm_compute::CLArithmeticAddition::validate(&aclInput0Info, - &aclInput1Info, - &aclOutputInfo, - g_AclConvertPolicy); - - return aclStatus; -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClAdditionWorkload.hpp b/src/backends/cl/workloads/ClAdditionWorkload.hpp deleted file mode 100644 index 62bd0ae20b..0000000000 --- a/src/backends/cl/workloads/ClAdditionWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLElementwiseOperations.h> - -namespace armnn -{ - -class ClAdditionWorkload : public BaseWorkload<AdditionQueueDescriptor> -{ -public: - ClAdditionWorkload(const AdditionQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable arm_compute::CLArithmeticAddition m_Layer; -}; - -arm_compute::Status ClAdditionValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output); -} //namespace armnn diff --git a/src/backends/cl/workloads/ClArgMinMaxWorkload.cpp b/src/backends/cl/workloads/ClArgMinMaxWorkload.cpp deleted file mode 100644 index a79a7b286d..0000000000 --- a/src/backends/cl/workloads/ClArgMinMaxWorkload.cpp +++ /dev/null @@ -1,82 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClArgMinMaxWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <armnnUtils/TensorUtils.hpp> - -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -namespace -{ -unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int axisIndex) -{ - return (numDimensions - axisIndex) - 1; -} - -} //namespace - -namespace armnn -{ - -arm_compute::Status ClArgMinMaxWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const ArgMinMaxDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - auto numDims = input.GetNumDimensions(); - auto unsignedAxis = armnnUtils::GetUnsignedAxis(numDims, descriptor.m_Axis); - int aclAxis = boost::numeric_cast<int>(CalcAclAxis(numDims, unsignedAxis)); - - if (descriptor.m_Function == ArgMinMaxFunction::Max) - { - return arm_compute::CLArgMinMaxLayer::validate(&aclInput, aclAxis, &aclOutput, - arm_compute::ReductionOperation::ARG_IDX_MAX); - } - else - { - return arm_compute::CLArgMinMaxLayer::validate(&aclInput, aclAxis, &aclOutput, - arm_compute::ReductionOperation::ARG_IDX_MIN); - } -} - - -ClArgMinMaxWorkload::ClArgMinMaxWorkload(const ArgMinMaxQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<ArgMinMaxQueueDescriptor>(descriptor, info) -{ - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor(); - - auto numDims = info.m_InputTensorInfos[0].GetNumDimensions(); - auto unsignedAxis = armnnUtils::GetUnsignedAxis(numDims, m_Data.m_Parameters.m_Axis); - int aclAxis = boost::numeric_cast<int>(CalcAclAxis(numDims, unsignedAxis)); - - if (m_Data.m_Parameters.m_Function == ArgMinMaxFunction::Max) - { - m_ArgMinMaxLayer.configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MAX); - } - else - { - m_ArgMinMaxLayer.configure(&input, aclAxis, &output, arm_compute::ReductionOperation::ARG_IDX_MIN); - } -} - -void ClArgMinMaxWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClArgMinMaxWorkload_Execute"); - RunClFunction(m_ArgMinMaxLayer, CHECK_LOCATION()); -} - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClArgMinMaxWorkload.hpp b/src/backends/cl/workloads/ClArgMinMaxWorkload.hpp deleted file mode 100644 index 54f28e6175..0000000000 --- a/src/backends/cl/workloads/ClArgMinMaxWorkload.hpp +++ /dev/null @@ -1,30 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/core/Error.h> -#include <arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h> - -namespace armnn -{ - -arm_compute::Status ClArgMinMaxWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const ArgMinMaxDescriptor& descriptor); - -class ClArgMinMaxWorkload : public BaseWorkload<ArgMinMaxQueueDescriptor> -{ -public: - ClArgMinMaxWorkload(const ArgMinMaxQueueDescriptor& descriptor, const WorkloadInfo& info); - virtual void Execute() const override; - -private: - mutable arm_compute::CLArgMinMaxLayer m_ArgMinMaxLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp b/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp deleted file mode 100644 index fa0be85100..0000000000 --- a/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.cpp +++ /dev/null @@ -1,106 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClBatchNormalizationFloatWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <cl/ClLayerSupport.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClBatchNormalizationValidate(const TensorInfo& input, - const TensorInfo& output, - const TensorInfo& mean, - const TensorInfo& var, - const TensorInfo& beta, - const TensorInfo& gamma, - const BatchNormalizationDescriptor &desc) -{ - const arm_compute::TensorInfo aclInputInfo = - armcomputetensorutils::BuildArmComputeTensorInfo(input, desc.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = - armcomputetensorutils::BuildArmComputeTensorInfo(output, desc.m_DataLayout); - const arm_compute::TensorInfo aclMeanInfo = - armcomputetensorutils::BuildArmComputeTensorInfo(mean, desc.m_DataLayout); - const arm_compute::TensorInfo aclVarInfo = - armcomputetensorutils::BuildArmComputeTensorInfo(var, desc.m_DataLayout); - const arm_compute::TensorInfo aclBetaInfo = - armcomputetensorutils::BuildArmComputeTensorInfo(beta, desc.m_DataLayout); - const arm_compute::TensorInfo aclGammaInfo = - armcomputetensorutils::BuildArmComputeTensorInfo(gamma, desc.m_DataLayout); - - return arm_compute::CLBatchNormalizationLayer::validate(&aclInputInfo, - &aclOutputInfo, - &aclMeanInfo, - &aclVarInfo, - &aclBetaInfo, - &aclGammaInfo, - desc.m_Eps); -} - -ClBatchNormalizationFloatWorkload::ClBatchNormalizationFloatWorkload( - const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info) - : FloatWorkload<BatchNormalizationQueueDescriptor>(descriptor, info) -{ - m_Mean = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_Mean, m_Data.m_Mean->GetTensorInfo()); - - m_Variance = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_Variance, m_Data.m_Variance->GetTensorInfo()); - - m_Gamma = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_Gamma, m_Data.m_Gamma->GetTensorInfo()); - - m_Beta = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_Beta, m_Data.m_Beta->GetTensorInfo()); - - m_Data.ValidateInputsOutputs("ClBatchNormalizationFloatWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - m_Layer.configure(&input, - &output, - m_Mean.get(), - m_Variance.get(), - m_Beta.get(), - m_Gamma.get(), - m_Data.m_Parameters.m_Eps); - - InitializeArmComputeClTensorData(*m_Mean, m_Data.m_Mean); - InitializeArmComputeClTensorData(*m_Variance, m_Data.m_Variance); - InitializeArmComputeClTensorData(*m_Beta, m_Data.m_Beta); - InitializeArmComputeClTensorData(*m_Gamma, m_Data.m_Gamma); - - // Force Compute Library to perform the necessary copying and reshaping, after which - // delete all the input tensors that will no longer be needed - m_Layer.prepare(); - FreeUnusedTensors(); -} - -void ClBatchNormalizationFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClBatchNormalizationFloatWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -void ClBatchNormalizationFloatWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_Mean); - FreeTensorIfUnused(m_Variance); - FreeTensorIfUnused(m_Gamma); - FreeTensorIfUnused(m_Beta); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.hpp b/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.hpp deleted file mode 100644 index e94bef20ac..0000000000 --- a/src/backends/cl/workloads/ClBatchNormalizationFloatWorkload.hpp +++ /dev/null @@ -1,47 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/CLTensor.h> -#include <arm_compute/runtime/CL/functions/CLBatchNormalizationLayer.h> - -namespace armnn -{ - -arm_compute::Status ClBatchNormalizationValidate(const TensorInfo& input, - const TensorInfo& output, - const TensorInfo& mean, - const TensorInfo& var, - const TensorInfo& beta, - const TensorInfo& gamma, - const BatchNormalizationDescriptor& desc); - -class ClBatchNormalizationFloatWorkload : public FloatWorkload<BatchNormalizationQueueDescriptor> -{ -public: - ClBatchNormalizationFloatWorkload(const BatchNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info); - - using FloatWorkload<BatchNormalizationQueueDescriptor>::FloatWorkload; - void Execute() const override; - -private: - mutable arm_compute::CLBatchNormalizationLayer m_Layer; - - std::unique_ptr<arm_compute::CLTensor> m_Mean; - std::unique_ptr<arm_compute::CLTensor> m_Variance; - std::unique_ptr<arm_compute::CLTensor> m_Gamma; - std::unique_ptr<arm_compute::CLTensor> m_Beta; - - void FreeUnusedTensors(); -}; - -} //namespace armnn - - - - diff --git a/src/backends/cl/workloads/ClBatchToSpaceNdWorkload.cpp b/src/backends/cl/workloads/ClBatchToSpaceNdWorkload.cpp deleted file mode 100644 index a714e031e4..0000000000 --- a/src/backends/cl/workloads/ClBatchToSpaceNdWorkload.cpp +++ /dev/null @@ -1,64 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClBatchToSpaceNdWorkload.hpp" - -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -ClBatchToSpaceNdWorkload::ClBatchToSpaceNdWorkload(const BatchToSpaceNdQueueDescriptor& desc, - const WorkloadInfo& info) - : BaseWorkload<BatchToSpaceNdQueueDescriptor>(desc, info) -{ - m_Data.ValidateInputsOutputs("ClBatchToSpaceNdWorkload", 1, 1); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - input.info()->set_data_layout(aclDataLayout); - - // ArmNN blockShape is [H, W] Cl asks for W, H - int32_t blockHeight = boost::numeric_cast<int32_t>(desc.m_Parameters.m_BlockShape[0]); - int32_t blockWidth = boost::numeric_cast<int32_t>(desc.m_Parameters.m_BlockShape[1]); - - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - output.info()->set_data_layout(aclDataLayout); - - m_Layer.configure(&input, blockWidth, blockHeight, &output); -} - -void ClBatchToSpaceNdWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClBatchToSpaceNdWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -arm_compute::Status ClBatchToSpaceNdWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const BatchToSpaceNdDescriptor& desc) { - DataLayout dataLayout = desc.m_DataLayout; - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, dataLayout); - - // ArmNN blockShape is [H, W] Cl asks for W, H - int32_t blockHeight = boost::numeric_cast<int32_t>(desc.m_BlockShape[0]); - int32_t blockWidth = boost::numeric_cast<int32_t>(desc.m_BlockShape[1]); - - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, dataLayout); - - const arm_compute::Status aclStatus = arm_compute::CLBatchToSpaceLayer::validate(&aclInputInfo, - blockWidth, - blockHeight, - &aclOutputInfo); - return aclStatus; -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClBatchToSpaceNdWorkload.hpp b/src/backends/cl/workloads/ClBatchToSpaceNdWorkload.hpp deleted file mode 100644 index 881b294097..0000000000 --- a/src/backends/cl/workloads/ClBatchToSpaceNdWorkload.hpp +++ /dev/null @@ -1,30 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> -#include <arm_compute/runtime/CL/functions/CLBatchToSpaceLayer.h> - -namespace armnn -{ - -arm_compute::Status ClBatchToSpaceNdWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const BatchToSpaceNdDescriptor& desc); - -class ClBatchToSpaceNdWorkload : public BaseWorkload<BatchToSpaceNdQueueDescriptor> -{ -public: - ClBatchToSpaceNdWorkload(const BatchToSpaceNdQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - - mutable arm_compute::CLBatchToSpaceLayer m_Layer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConcatWorkload.cpp b/src/backends/cl/workloads/ClConcatWorkload.cpp deleted file mode 100644 index 5370466163..0000000000 --- a/src/backends/cl/workloads/ClConcatWorkload.cpp +++ /dev/null @@ -1,102 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// -#include "ClConcatWorkload.hpp" -#include "ClWorkloadUtils.hpp" -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -#include <arm_compute/core/Types.h> -#include <arm_compute/runtime/CL/functions/CLConcatenateLayer.h> - -#include <boost/polymorphic_pointer_cast.hpp> - -namespace armnn -{ -using namespace armcomputetensorutils; - -namespace -{ -size_t CalcAxis(const OriginsDescriptor& desc) -{ - return (desc.GetNumDimensions() - desc.GetConcatAxis()) - 1; -} -} //namespace - -arm_compute::Status ClConcatWorkloadValidate(const std::vector<const TensorInfo*>& inputs, - const TensorInfo& output, - const OriginsDescriptor& descriptor) -{ - std::vector<arm_compute::TensorInfo> aclInputs; - for (const TensorInfo* input : inputs) - { - arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(*input, armnn::DataLayout::NCHW); - aclInputs.emplace_back(aclInputInfo); - } - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - std::vector<arm_compute::ITensorInfo*> aclInputPtrs; - for (arm_compute::ITensorInfo& input : aclInputs) - { - aclInputPtrs.emplace_back(&input); - } - - size_t aclAxis = CalcAxis(descriptor); - return arm_compute::CLConcatenateLayer::validate(aclInputPtrs, &aclOutputInfo, aclAxis); -} - -ClConcatWorkload::ClConcatWorkload(const ConcatQueueDescriptor& descriptor, const WorkloadInfo& info) -: BaseWorkload<ConcatQueueDescriptor>(descriptor, info) -{ - bool allInputsAreSubtensors = true; - - // Check that all inputs are sub-tensors - for (auto input : descriptor.m_Inputs) - { - if (!input->GetParent()) - { - // Non sub-tensor input found so we need to execute the concat function - allInputsAreSubtensors = false; - break; - } - } - - if (allInputsAreSubtensors) - { - // Can skip configuring the concat function since it's not executed - return; - } - - std::vector<arm_compute::ICLTensor *> aclInputs; - for (auto input : m_Data.m_Inputs) - { - arm_compute::ICLTensor& aclInput = boost::polymorphic_pointer_downcast<IClTensorHandle>(input)->GetTensor(); - aclInputs.emplace_back(&aclInput); - } - arm_compute::ICLTensor& output = boost::polymorphic_pointer_downcast<IClTensorHandle>( - m_Data.m_Outputs[0])->GetTensor(); - - // Create the layer function - auto layer = std::make_unique<arm_compute::CLConcatenateLayer>(); - - // Configure input and output tensors - size_t aclAxis = CalcAxis(descriptor.m_Parameters); - layer->configure(aclInputs, &output, aclAxis); - - // Prepare - layer->prepare(); - m_Layer = std::move(layer); -} - -void ClConcatWorkload::Execute() const -{ - if (m_Layer) - { - ARMNN_SCOPED_PROFILING_EVENT_CL("ClConcatWorkload_Execute"); - m_Layer->run(); - } -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConcatWorkload.hpp b/src/backends/cl/workloads/ClConcatWorkload.hpp deleted file mode 100644 index 772bc094ea..0000000000 --- a/src/backends/cl/workloads/ClConcatWorkload.hpp +++ /dev/null @@ -1,31 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/core/Error.h> -#include <arm_compute/runtime/IFunction.h> - -namespace armnn -{ - -arm_compute::Status ClConcatWorkloadValidate(const std::vector<const TensorInfo*>& inputs, - const TensorInfo& output, - const OriginsDescriptor& descriptor); - -class ClConcatWorkload : public BaseWorkload<ConcatQueueDescriptor> -{ -public: - ClConcatWorkload(const ConcatQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable std::unique_ptr<arm_compute::IFunction> m_Layer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConstantWorkload.cpp b/src/backends/cl/workloads/ClConstantWorkload.cpp deleted file mode 100644 index 39ae14eaf3..0000000000 --- a/src/backends/cl/workloads/ClConstantWorkload.cpp +++ /dev/null @@ -1,68 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClConstantWorkload.hpp" - -#include <Half.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ - -ClConstantWorkload::ClConstantWorkload(const ConstantQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<ConstantQueueDescriptor>(descriptor, info) - , m_RanOnce(false) -{ -} - -void ClConstantWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClConstantWorkload_Execute"); - - // The intermediate tensor held by the corresponding layer output handler can be initialised with the given data - // on the first inference, then reused for subsequent inferences. - // The initialisation cannot happen at workload construction time since the ACL kernel for the next layer may not - // have been configured at the time. - if (!m_RanOnce) - { - const ConstantQueueDescriptor& data = this->m_Data; - - BOOST_ASSERT(data.m_LayerOutput != nullptr); - arm_compute::CLTensor& output = static_cast<ClTensorHandle*>(data.m_Outputs[0])->GetTensor(); - arm_compute::DataType computeDataType = static_cast<ClTensorHandle*>(data.m_Outputs[0])->GetDataType(); - - switch (computeDataType) - { - case arm_compute::DataType::F16: - { - CopyArmComputeClTensorData(output, data.m_LayerOutput->GetConstTensor<Half>()); - break; - } - case arm_compute::DataType::F32: - { - CopyArmComputeClTensorData(output, data.m_LayerOutput->GetConstTensor<float>()); - break; - } - case arm_compute::DataType::QASYMM8: - { - CopyArmComputeClTensorData(output, data.m_LayerOutput->GetConstTensor<uint8_t>()); - break; - } - default: - { - BOOST_ASSERT_MSG(false, "Unknown data type"); - break; - } - } - - m_RanOnce = true; - } -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConstantWorkload.hpp b/src/backends/cl/workloads/ClConstantWorkload.hpp deleted file mode 100644 index 75325dc409..0000000000 --- a/src/backends/cl/workloads/ClConstantWorkload.hpp +++ /dev/null @@ -1,23 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -namespace armnn -{ -class ClConstantWorkload : public BaseWorkload<ConstantQueueDescriptor> -{ -public: - ClConstantWorkload(const ConstantQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable bool m_RanOnce; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConvertFp16ToFp32Workload.cpp b/src/backends/cl/workloads/ClConvertFp16ToFp32Workload.cpp deleted file mode 100644 index d2e86f8c94..0000000000 --- a/src/backends/cl/workloads/ClConvertFp16ToFp32Workload.cpp +++ /dev/null @@ -1,56 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClConvertFp16ToFp32Workload.hpp" -#include <cl/ClTensorHandle.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE; - -ClConvertFp16ToFp32Workload::ClConvertFp16ToFp32Workload( - const ConvertFp16ToFp32QueueDescriptor& descriptor, const WorkloadInfo& info) : - Float16ToFloat32Workload<ConvertFp16ToFp32QueueDescriptor>(descriptor, info) -{ - this->m_Data.ValidateInputsOutputs("ClConvertFp16ToFp32Workload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor(); - - m_Layer.configure(&input, &output, g_AclConvertPolicy, 0); -} - -void ClConvertFp16ToFp32Workload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvertFp16ToFp32Workload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -arm_compute::Status ClConvertFp16ToFp32WorkloadValidate(const TensorInfo& input, const TensorInfo& output) -{ - if (input.GetDataType() != DataType::Float16) - { - return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, "Input should be Float16"); - } - if (output.GetDataType() != DataType::Float32) - { - return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, "Output should be Float32"); - } - - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - const arm_compute::Status aclStatus = arm_compute::CLDepthConvertLayer::validate( - &aclInputInfo, &aclOutputInfo, g_AclConvertPolicy, 0); - - return aclStatus; -} - - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConvertFp16ToFp32Workload.hpp b/src/backends/cl/workloads/ClConvertFp16ToFp32Workload.hpp deleted file mode 100644 index ef5c9b6497..0000000000 --- a/src/backends/cl/workloads/ClConvertFp16ToFp32Workload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLDepthConvertLayer.h> - -namespace armnn -{ - -class ClConvertFp16ToFp32Workload : public Float16ToFloat32Workload<ConvertFp16ToFp32QueueDescriptor> -{ -public: - - ClConvertFp16ToFp32Workload(const ConvertFp16ToFp32QueueDescriptor& descriptor, const WorkloadInfo& info); - virtual void Execute() const override; - -private: - mutable arm_compute::CLDepthConvertLayer m_Layer; -}; - -arm_compute::Status ClConvertFp16ToFp32WorkloadValidate(const TensorInfo& input, const TensorInfo& output); - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConvertFp32ToFp16Workload.cpp b/src/backends/cl/workloads/ClConvertFp32ToFp16Workload.cpp deleted file mode 100644 index 3f528a1532..0000000000 --- a/src/backends/cl/workloads/ClConvertFp32ToFp16Workload.cpp +++ /dev/null @@ -1,56 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClConvertFp32ToFp16Workload.hpp" -#include <cl/ClTensorHandle.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE; - -ClConvertFp32ToFp16Workload::ClConvertFp32ToFp16Workload( - const ConvertFp32ToFp16QueueDescriptor& descriptor, const WorkloadInfo& info) : - Float32ToFloat16Workload<ConvertFp32ToFp16QueueDescriptor>(descriptor, info) -{ - this->m_Data.ValidateInputsOutputs("ClConvertFp32ToFp16Workload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor(); - - m_Layer.configure(&input, &output, g_AclConvertPolicy, 0); -} - -void ClConvertFp32ToFp16Workload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvertFp32ToFp16Workload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -arm_compute::Status ClConvertFp32ToFp16WorkloadValidate(const TensorInfo& input, const TensorInfo& output) -{ - if (input.GetDataType() != DataType::Float32) - { - return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, "Input should be Float32"); - } - if (output.GetDataType() != DataType::Float16) - { - return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR, "Output should be Float16"); - } - - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - const arm_compute::Status aclStatus = arm_compute::CLDepthConvertLayer::validate( - &aclInputInfo, &aclOutputInfo, g_AclConvertPolicy, 0); - - return aclStatus; -} - - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConvertFp32ToFp16Workload.hpp b/src/backends/cl/workloads/ClConvertFp32ToFp16Workload.hpp deleted file mode 100644 index 6e04e39425..0000000000 --- a/src/backends/cl/workloads/ClConvertFp32ToFp16Workload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLDepthConvertLayer.h> - -namespace armnn -{ - -class ClConvertFp32ToFp16Workload : public Float32ToFloat16Workload<ConvertFp32ToFp16QueueDescriptor> -{ -public: - - ClConvertFp32ToFp16Workload(const ConvertFp32ToFp16QueueDescriptor& descriptor, const WorkloadInfo& info); - virtual void Execute() const override; - -private: - mutable arm_compute::CLDepthConvertLayer m_Layer; -}; - -arm_compute::Status ClConvertFp32ToFp16WorkloadValidate(const TensorInfo& input, const TensorInfo& output); - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp deleted file mode 100644 index e8af0ee3b7..0000000000 --- a/src/backends/cl/workloads/ClConvolution2dWorkload.cpp +++ /dev/null @@ -1,122 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClConvolution2dWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h> - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const Convolution2dDescriptor& descriptor, - const TensorInfo& weights, - const Optional<TensorInfo>& biases) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); - - const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX, - descriptor.m_DilationY); - - arm_compute::TensorInfo aclBiasesInfo; - arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; - - if (descriptor.m_BiasEnabled) - { - BOOST_ASSERT(biases.has_value()); - - aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); - optionalAclBiasesInfo = &aclBiasesInfo; - } - - arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor); - - return arm_compute::CLConvolutionLayer::validate(&aclInputInfo, - &aclWeightsInfo, - optionalAclBiasesInfo, - &aclOutputInfo, - layerInfo, - arm_compute::WeightsInfo(), - aclDilationInfo); -} - -ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) - : BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info) - , m_ConvolutionLayer(memoryManager) -{ - // todo: check tensor shapes match. - const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo(); - - m_KernelTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout); - - const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(m_Data.m_Parameters.m_DilationX, - m_Data.m_Parameters.m_DilationY); - - if (m_Data.m_Parameters.m_BiasEnabled) - { - m_BiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); - } - - m_Data.ValidateInputsOutputs("ClConvolution2dWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters); - - m_ConvolutionLayer.configure(&input, - m_KernelTensor.get(), - m_BiasTensor.get(), - &output, - padStrideInfo, - arm_compute::WeightsInfo(), - aclDilationInfo); - - InitializeArmComputeClTensorData(*m_KernelTensor, m_Data.m_Weight); - - if (m_BiasTensor) - { - InitializeArmComputeClTensorData(*m_BiasTensor, m_Data.m_Bias); - } - - // Force Compute Library to perform the necessary copying and reshaping, after which - // delete all the input tensors that will no longer be needed - m_ConvolutionLayer.prepare(); - FreeUnusedTensors(); -} - -void ClConvolution2dWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvolution2dWorkload_Execute"); - RunClFunction(m_ConvolutionLayer, CHECK_LOCATION()); -} - -void ClConvolution2dWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_KernelTensor); - FreeTensorIfUnused(m_BiasTensor); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.hpp b/src/backends/cl/workloads/ClConvolution2dWorkload.hpp deleted file mode 100644 index 6d7e9f3ea1..0000000000 --- a/src/backends/cl/workloads/ClConvolution2dWorkload.hpp +++ /dev/null @@ -1,44 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Tensor.hpp> -#include <armnn/Descriptors.hpp> - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h> -#include <arm_compute/runtime/MemoryManagerOnDemand.h> - -#include <memory> - -namespace armnn -{ - -arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const Convolution2dDescriptor& descriptor, - const TensorInfo& weights, - const Optional<TensorInfo>& biases); - -class ClConvolution2dWorkload : public BaseWorkload<Convolution2dQueueDescriptor> -{ -public: - ClConvolution2dWorkload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); - void Execute() const override; - -private: - mutable arm_compute::CLConvolutionLayer m_ConvolutionLayer; - - std::unique_ptr<arm_compute::CLTensor> m_KernelTensor; - std::unique_ptr<arm_compute::CLTensor> m_BiasTensor; - - void FreeUnusedTensors(); -}; - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClDepthToSpaceWorkload.cpp b/src/backends/cl/workloads/ClDepthToSpaceWorkload.cpp deleted file mode 100644 index 800a98409e..0000000000 --- a/src/backends/cl/workloads/ClDepthToSpaceWorkload.cpp +++ /dev/null @@ -1,66 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClDepthToSpaceWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <cl/ClTensorHandle.hpp> - -#include <boost/numeric/conversion/cast.hpp> -#include <boost/polymorphic_pointer_cast.hpp> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClDepthToSpaceWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const DepthToSpaceDescriptor& desc) -{ - DataLayout dataLayout = desc.m_DataLayout; - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, dataLayout); - - int32_t blockSize = boost::numeric_cast<int32_t>(desc.m_BlockSize); - - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, dataLayout); - - const arm_compute::Status aclStatus = arm_compute::CLDepthToSpaceLayer::validate(&aclInputInfo, - &aclOutputInfo, - blockSize); - return aclStatus; -} - -ClDepthToSpaceWorkload::ClDepthToSpaceWorkload(const DepthToSpaceQueueDescriptor& desc, - const WorkloadInfo& info) - : BaseWorkload<DepthToSpaceQueueDescriptor>(desc, info) -{ - m_Data.ValidateInputsOutputs("ClDepthToSpaceWorkload", 1, 1); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - - arm_compute::ICLTensor& input = - boost::polymorphic_pointer_downcast<IClTensorHandle>(m_Data.m_Inputs[0])->GetTensor(); - input.info()->set_data_layout(aclDataLayout); - - int32_t blockSize = boost::numeric_cast<int32_t>(desc.m_Parameters.m_BlockSize); - - arm_compute::ICLTensor& output = - boost::polymorphic_pointer_downcast<IClTensorHandle>(m_Data.m_Outputs[0])->GetTensor(); - output.info()->set_data_layout(aclDataLayout); - - m_Layer.configure(&input, &output, blockSize); -} - -void ClDepthToSpaceWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClDepthToSpaceWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClDepthToSpaceWorkload.hpp b/src/backends/cl/workloads/ClDepthToSpaceWorkload.hpp deleted file mode 100644 index de8b496669..0000000000 --- a/src/backends/cl/workloads/ClDepthToSpaceWorkload.hpp +++ /dev/null @@ -1,31 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Descriptors.hpp> - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLDepthToSpaceLayer.h> - -namespace armnn -{ - -arm_compute::Status ClDepthToSpaceWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const DepthToSpaceDescriptor& desc); - -class ClDepthToSpaceWorkload : public BaseWorkload<DepthToSpaceQueueDescriptor> -{ -public: - ClDepthToSpaceWorkload(const DepthToSpaceQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLDepthToSpaceLayer m_Layer; -}; - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp deleted file mode 100644 index 858eab4e00..0000000000 --- a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.cpp +++ /dev/null @@ -1,156 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClDepthwiseConvolutionWorkload.hpp" - -#include <ResolveType.hpp> -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <backendsCommon/WorkloadUtils.hpp> - -#include <arm_compute/runtime/CL/functions/CLDepthwiseConvolutionLayer.h> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClDepthwiseConvolutionWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const DepthwiseConvolution2dDescriptor& descriptor, - const TensorInfo& weights, - const Optional<TensorInfo>& biases) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - - // ArmNN's weight format is [ M, I, H, W ] - const unsigned int aclDepthMultiplier = weights.GetShape()[0]; - - // Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either - // [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library - TensorInfo weightsPermuted = ConvertWeightTensorInfoFromArmnnToAcl(weights, descriptor.m_DataLayout); - - // Convert the weights into the compute library format - const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weightsPermuted, descriptor.m_DataLayout); - - arm_compute::TensorInfo aclBiasesInfo; - arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; - - if (descriptor.m_BiasEnabled) - { - BOOST_ASSERT(biases.has_value()); - - aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); - optionalAclBiasesInfo = &aclBiasesInfo; - } - - const arm_compute::PadStrideInfo aclPadStrideInfo = BuildArmComputePadStrideInfo(descriptor); - const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D( - descriptor.m_DilationX, - descriptor.m_DilationY); - - return arm_compute::CLDepthwiseConvolutionLayer::validate(&aclInputInfo, - &aclWeightsInfo, - optionalAclBiasesInfo, - &aclOutputInfo, - aclPadStrideInfo, - aclDepthMultiplier, - arm_compute::ActivationLayerInfo(), - aclDilationInfo); - -} - -ClDepthwiseConvolutionWorkload::ClDepthwiseConvolutionWorkload( - const DepthwiseConvolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<DepthwiseConvolution2dQueueDescriptor>(descriptor, info) -{ - // Allocate a buffer for the swizzling of the weight tensor - std::unique_ptr<unsigned char[]> permuteBuffer(new unsigned char[m_Data.m_Weight->GetTensorInfo().GetNumBytes()]); - - // Convert the weight format from ArmNN's [ M, I, H, W ] (does NOT depend on the data layout) to either - // [ 1, H, W, I * M ] (if NHWC) or [ 1, I * M, H, W ] (if NCHW), as required by the compute library - ConstTensor weightPermuted = ConvertWeightTensorFromArmnnToAcl(m_Data.m_Weight, - m_Data.m_Parameters.m_DataLayout, - permuteBuffer.get()); - - // Convert the weights into the compute library format - m_KernelTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_KernelTensor, weightPermuted.GetInfo(), m_Data.m_Parameters.m_DataLayout); - - if (m_Data.m_Parameters.m_BiasEnabled) - { - m_BiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); - } - - const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D( - m_Data.m_Parameters.m_DilationX, - m_Data.m_Parameters.m_DilationY); - - - std::string name = std::string("ClDepthwiseConvolutionWorkload"); - m_Data.ValidateInputsOutputs(name, 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - // ArmNN's weight format is [ M, I, H, W ] - auto& weightInfo = m_Data.m_Weight->GetTensorInfo(); - - // Get the depth multiplier - const unsigned int depthMultiplier = weightInfo.GetShape()[0]; - - arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters); - - m_DepthwiseConvolutionLayer = std::make_unique<arm_compute::CLDepthwiseConvolutionLayer>(); - static_cast<arm_compute::CLDepthwiseConvolutionLayer*>(m_DepthwiseConvolutionLayer.get())->configure( - &input, - m_KernelTensor.get(), - m_BiasTensor.get(), - &output, - padStrideInfo, - depthMultiplier, - arm_compute::ActivationLayerInfo(), - aclDilationInfo); - - BOOST_ASSERT(m_DepthwiseConvolutionLayer); - - ScopedCpuTensorHandle weightsPermutedHandle(weightPermuted); - InitializeArmComputeClTensorData(*m_KernelTensor, &weightsPermutedHandle); - - if (m_BiasTensor) - { - InitializeArmComputeClTensorData(*m_BiasTensor, m_Data.m_Bias); - } - - m_DepthwiseConvolutionLayer->prepare(); - FreeUnusedTensors(); -} - -void ClDepthwiseConvolutionWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_KernelTensor); - FreeTensorIfUnused(m_BiasTensor); -} - -void ClDepthwiseConvolutionWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClDepthwiseConvolutionWorkload_Execute"); - BOOST_ASSERT(m_DepthwiseConvolutionLayer); - - RunClFunction(*m_DepthwiseConvolutionLayer, CHECK_LOCATION()); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.hpp b/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.hpp deleted file mode 100644 index fc277b9947..0000000000 --- a/src/backends/cl/workloads/ClDepthwiseConvolutionWorkload.hpp +++ /dev/null @@ -1,42 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/IFunction.h> -#include <arm_compute/core/Error.h> -#include <arm_compute/runtime/CL/CLTensor.h> - -namespace armnn -{ - -arm_compute::Status ClDepthwiseConvolutionWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const DepthwiseConvolution2dDescriptor& descriptor, - const TensorInfo& weights, - const Optional<TensorInfo>& biases); - -class ClDepthwiseConvolutionWorkload : public BaseWorkload<DepthwiseConvolution2dQueueDescriptor> -{ -public: - using BaseWorkload<DepthwiseConvolution2dQueueDescriptor>::m_Data; - - ClDepthwiseConvolutionWorkload(const DepthwiseConvolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info); - - void Execute() const override; - -protected: - std::unique_ptr<arm_compute::IFunction> m_DepthwiseConvolutionLayer; - - std::unique_ptr<arm_compute::CLTensor> m_KernelTensor; - std::unique_ptr<arm_compute::CLTensor> m_BiasTensor; - - void FreeUnusedTensors(); -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClDequantizeWorkload.cpp b/src/backends/cl/workloads/ClDequantizeWorkload.cpp deleted file mode 100644 index eca795de7e..0000000000 --- a/src/backends/cl/workloads/ClDequantizeWorkload.cpp +++ /dev/null @@ -1,57 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClDequantizeWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <arm_compute/core/Types.h> - -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> - -#include <boost/polymorphic_pointer_cast.hpp> - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClDequantizeWorkloadValidate(const TensorInfo& input, const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - return arm_compute::CLDequantizationLayer::validate(&aclInputInfo, &aclOutputInfo); -} - -ClDequantizeWorkload::ClDequantizeWorkload(const DequantizeQueueDescriptor& descriptor, - const WorkloadInfo& workloadInfo) - : BaseWorkload<DequantizeQueueDescriptor>(descriptor, workloadInfo) -{ - m_Data.ValidateInputsOutputs("ClDequantizeWorkload", 1, 1); - - arm_compute::ICLTensor& input = boost::polymorphic_pointer_downcast<IClTensorHandle>( - m_Data.m_Inputs[0])->GetTensor(); - - arm_compute::ICLTensor& output = boost::polymorphic_pointer_downcast<IClTensorHandle>( - m_Data.m_Outputs[0])->GetTensor(); - - m_Layer.reset(new arm_compute::CLDequantizationLayer()); - m_Layer->configure(&input, &output); - m_Layer->prepare(); -} - -void ClDequantizeWorkload::Execute() const -{ - if (m_Layer) - { - ARMNN_SCOPED_PROFILING_EVENT_CL("ClDequantizeWorkload_Execute"); - m_Layer->run(); - } -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClDequantizeWorkload.hpp b/src/backends/cl/workloads/ClDequantizeWorkload.hpp deleted file mode 100644 index 6e61da2ebf..0000000000 --- a/src/backends/cl/workloads/ClDequantizeWorkload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLDequantizationLayer.h> - -namespace armnn -{ - -arm_compute::Status ClDequantizeWorkloadValidate(const TensorInfo& input, const TensorInfo& output); - -class ClDequantizeWorkload : public BaseWorkload<DequantizeQueueDescriptor> -{ -public: - ClDequantizeWorkload(const DequantizeQueueDescriptor& descriptor, const WorkloadInfo& workloadInfo); - - void Execute() const override; - -private: - mutable std::unique_ptr<arm_compute::CLDequantizationLayer> m_Layer; -}; - -} // namespace armnn
\ No newline at end of file diff --git a/src/backends/cl/workloads/ClDivisionFloatWorkload.cpp b/src/backends/cl/workloads/ClDivisionFloatWorkload.cpp deleted file mode 100644 index 2a27f8a9bc..0000000000 --- a/src/backends/cl/workloads/ClDivisionFloatWorkload.cpp +++ /dev/null @@ -1,46 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClDivisionFloatWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ - -arm_compute::Status ClDivisionWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput1 = armcomputetensorutils::BuildArmComputeTensorInfo(input0); - const arm_compute::TensorInfo aclInput2 = armcomputetensorutils::BuildArmComputeTensorInfo(input1); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - return arm_compute::CLArithmeticDivision::validate(&aclInput1, &aclInput2, &aclOutput); -} - - -ClDivisionFloatWorkload::ClDivisionFloatWorkload(const DivisionQueueDescriptor& descriptor, - const WorkloadInfo& info) - : FloatWorkload<DivisionQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClDivisionFloatWorkload", 2, 1); - - arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - // Construct - m_ArithmeticDivision.configure(&input0, &input1, &output); -} - -void ClDivisionFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClDivisionFloatWorkload_Execute"); - RunClFunction(m_ArithmeticDivision, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClDivisionFloatWorkload.hpp b/src/backends/cl/workloads/ClDivisionFloatWorkload.hpp deleted file mode 100644 index ddca87d78a..0000000000 --- a/src/backends/cl/workloads/ClDivisionFloatWorkload.hpp +++ /dev/null @@ -1,32 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLElementwiseOperations.h> - -namespace armnn -{ - -arm_compute::Status ClDivisionWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output); - -class ClDivisionFloatWorkload : public FloatWorkload<DivisionQueueDescriptor> -{ -public: - ClDivisionFloatWorkload(const DivisionQueueDescriptor& descriptor, const - WorkloadInfo& info); - - using FloatWorkload<DivisionQueueDescriptor>::FloatWorkload; - void Execute() const override; - -private: - mutable arm_compute::CLArithmeticDivision m_ArithmeticDivision; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClFloorFloatWorkload.cpp b/src/backends/cl/workloads/ClFloorFloatWorkload.cpp deleted file mode 100644 index f38342ed39..0000000000 --- a/src/backends/cl/workloads/ClFloorFloatWorkload.cpp +++ /dev/null @@ -1,40 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClFloorFloatWorkload.hpp" -#include <cl/ClTensorHandle.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ - -arm_compute::Status ClFloorWorkloadValidate(const TensorInfo& input, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - return arm_compute::CLFloor::validate(&aclInput, &aclOutput); -} - -ClFloorFloatWorkload::ClFloorFloatWorkload(const FloorQueueDescriptor& descriptor, const WorkloadInfo& info) - : FloatWorkload<FloorQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClFloorFloatWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_Layer.configure(&input, &output); -} - -void ClFloorFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClFloorFloatWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClFloorFloatWorkload.hpp b/src/backends/cl/workloads/ClFloorFloatWorkload.hpp deleted file mode 100644 index 1ddaddff0b..0000000000 --- a/src/backends/cl/workloads/ClFloorFloatWorkload.hpp +++ /dev/null @@ -1,33 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLFloor.h> - -namespace armnn -{ - -arm_compute::Status ClFloorWorkloadValidate(const TensorInfo& input, - const TensorInfo& output); - -class ClFloorFloatWorkload : public FloatWorkload<FloorQueueDescriptor> -{ -public: - ClFloorFloatWorkload(const FloorQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable arm_compute::CLFloor m_Layer; -}; - -} //namespace armnn - - - - diff --git a/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp b/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp deleted file mode 100644 index 60eb138b42..0000000000 --- a/src/backends/cl/workloads/ClFullyConnectedWorkload.cpp +++ /dev/null @@ -1,96 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClFullyConnectedWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <aclCommon/ArmComputeUtils.hpp> -#include <cl/ClLayerSupport.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClFullyConnectedWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const TensorInfo& weights, - const TensorInfo& biases, - const FullyConnectedDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output); - const arm_compute::TensorInfo aclWeights = BuildArmComputeTensorInfo(weights); - - arm_compute::TensorInfo aclBiases; - arm_compute::TensorInfo *optionalAclBiases = nullptr; - if (descriptor.m_BiasEnabled) - { - aclBiases = BuildArmComputeTensorInfo(biases); - optionalAclBiases = &aclBiases; - } - - const arm_compute::FullyConnectedLayerInfo fullyConnectedLayerInfo = - ConvertFullyConnectedDescriptorToAclFullyConnectedLayerInfo(descriptor); - - return arm_compute::CLFullyConnectedLayer::validate(&aclInput, - &aclWeights, - optionalAclBiases, - &aclOutput, - fullyConnectedLayerInfo); -} - -ClFullyConnectedWorkload::ClFullyConnectedWorkload(const FullyConnectedQueueDescriptor& descriptor, - const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) - : BaseWorkload<FullyConnectedQueueDescriptor>(descriptor, info) - , m_FullyConnectedLayer(memoryManager) -{ - m_WeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_WeightsTensor, m_Data.m_Weight->GetTensorInfo()); - - if (m_Data.m_Parameters.m_BiasEnabled) - { - m_BiasesTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo()); - } - - m_Data.ValidateInputsOutputs("ClFullyConnectedWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - // Construct - arm_compute::FullyConnectedLayerInfo fc_info; - fc_info.transpose_weights = m_Data.m_Parameters.m_TransposeWeightMatrix; - m_FullyConnectedLayer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, fc_info); - - InitializeArmComputeClTensorData(*m_WeightsTensor, m_Data.m_Weight); - - if (m_BiasesTensor) - { - InitializeArmComputeClTensorData(*m_BiasesTensor, m_Data.m_Bias); - } - - // Force Compute Library to perform the necessary copying and reshaping, after which - // delete all the input tensors that will no longer be needed - m_FullyConnectedLayer.prepare(); - FreeUnusedTensors(); -} - -void ClFullyConnectedWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClFullyConnectedWorkload_Execute"); - RunClFunction(m_FullyConnectedLayer, CHECK_LOCATION()); -} - -void ClFullyConnectedWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_WeightsTensor); - FreeTensorIfUnused(m_BiasesTensor); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClFullyConnectedWorkload.hpp b/src/backends/cl/workloads/ClFullyConnectedWorkload.hpp deleted file mode 100644 index e13436eaa5..0000000000 --- a/src/backends/cl/workloads/ClFullyConnectedWorkload.hpp +++ /dev/null @@ -1,43 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLFullyConnectedLayer.h> -#include <arm_compute/runtime/MemoryManagerOnDemand.h> - -#include <memory> - -namespace armnn -{ - -arm_compute::Status ClFullyConnectedWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const TensorInfo& weights, - const TensorInfo& biases, - const FullyConnectedDescriptor& descriptor); - -class ClFullyConnectedWorkload : public armnn::BaseWorkload<armnn::FullyConnectedQueueDescriptor> -{ -public: - ClFullyConnectedWorkload(const armnn::FullyConnectedQueueDescriptor& descriptor, - const armnn::WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); - - using armnn::BaseWorkload<armnn::FullyConnectedQueueDescriptor>::m_Data; - void Execute() const override; - -private: - mutable arm_compute::CLFullyConnectedLayer m_FullyConnectedLayer; - - std::unique_ptr<arm_compute::CLTensor> m_WeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_BiasesTensor; - - void FreeUnusedTensors(); -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClGreaterWorkload.cpp b/src/backends/cl/workloads/ClGreaterWorkload.cpp deleted file mode 100644 index b086122bdc..0000000000 --- a/src/backends/cl/workloads/ClGreaterWorkload.cpp +++ /dev/null @@ -1,65 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClGreaterWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClGreaterWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0); - const arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - const arm_compute::Status aclStatus = arm_compute::CLComparison::validate( - &aclInput0Info, - &aclInput1Info, - &aclOutputInfo, - arm_compute::ComparisonOperation::Greater); - - return aclStatus; -} - -template<DataType T> -ClGreaterWorkload<T>::ClGreaterWorkload(const GreaterQueueDescriptor& descriptor, - const WorkloadInfo& info) - : MultiTypedWorkload<GreaterQueueDescriptor, T, DataType::Boolean>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClGreaterWorkload", 2, 1); - - arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_GreaterLayer.configure(&input0, &input1, &output, arm_compute::ComparisonOperation::Greater); -} - -template<DataType T> -void ClGreaterWorkload<T>::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClGreaterWorkload_Execute"); - RunClFunction(m_GreaterLayer, CHECK_LOCATION()); -} - -template class ClGreaterWorkload<DataType::Float32>; -template class ClGreaterWorkload<DataType::QAsymmU8>; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClGreaterWorkload.hpp b/src/backends/cl/workloads/ClGreaterWorkload.hpp deleted file mode 100644 index 862e168334..0000000000 --- a/src/backends/cl/workloads/ClGreaterWorkload.hpp +++ /dev/null @@ -1,34 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLComparison.h> - -namespace armnn -{ - -arm_compute::Status ClGreaterWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output); - -template<DataType T> -class ClGreaterWorkload : public MultiTypedWorkload<GreaterQueueDescriptor, T, DataType::Boolean> -{ -public: - ClGreaterWorkload(const GreaterQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - using MultiTypedWorkload<GreaterQueueDescriptor, T, DataType::Boolean>::m_Data; - mutable arm_compute::CLComparison m_GreaterLayer; -}; - -using ClGreaterFloat32Workload = ClGreaterWorkload<DataType::Float32>; -using ClGreaterUint8Workload = ClGreaterWorkload<DataType::QAsymmU8>; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClInstanceNormalizationWorkload.cpp b/src/backends/cl/workloads/ClInstanceNormalizationWorkload.cpp deleted file mode 100644 index 50cf345a7f..0000000000 --- a/src/backends/cl/workloads/ClInstanceNormalizationWorkload.cpp +++ /dev/null @@ -1,59 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClInstanceNormalizationWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <cl/ClTensorHandle.hpp> - -using namespace armnn::armcomputetensorutils; - -namespace armnn -{ - -arm_compute::Status ClInstanceNormalizationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const InstanceNormalizationDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - - return arm_compute::CLInstanceNormalizationLayer::validate(&aclInputInfo, - &aclOutputInfo, - descriptor.m_Gamma, - descriptor.m_Beta, - descriptor.m_Eps); -} - -ClInstanceNormalizationWorkload::ClInstanceNormalizationWorkload( - const InstanceNormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<InstanceNormalizationQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClInstanceNormalizationWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - m_Layer.configure(&input, - &output, - descriptor.m_Parameters.m_Gamma, - descriptor.m_Parameters.m_Beta, - descriptor.m_Parameters.m_Eps); -}; - -void ClInstanceNormalizationWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClInstanceNormalizationWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClInstanceNormalizationWorkload.hpp b/src/backends/cl/workloads/ClInstanceNormalizationWorkload.hpp deleted file mode 100644 index 0e37bdcc9b..0000000000 --- a/src/backends/cl/workloads/ClInstanceNormalizationWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLInstanceNormalizationLayer.h> - -namespace armnn -{ - -arm_compute::Status ClInstanceNormalizationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const InstanceNormalizationDescriptor& descriptor); - -class ClInstanceNormalizationWorkload : public BaseWorkload<InstanceNormalizationQueueDescriptor> -{ -public: - ClInstanceNormalizationWorkload(const InstanceNormalizationQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLInstanceNormalizationLayer m_Layer; -}; - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClL2NormalizationFloatWorkload.cpp b/src/backends/cl/workloads/ClL2NormalizationFloatWorkload.cpp deleted file mode 100644 index e1a61277d5..0000000000 --- a/src/backends/cl/workloads/ClL2NormalizationFloatWorkload.cpp +++ /dev/null @@ -1,53 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClL2NormalizationFloatWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeUtils.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClL2NormalizationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const L2NormalizationDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInput = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutput = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - - int axis = (descriptor.m_DataLayout == DataLayout::NCHW) ? 2 : 0; - - return arm_compute::CLL2NormalizeLayer::validate(&aclInput, &aclOutput, axis, descriptor.m_Eps); -} - -ClL2NormalizationFloatWorkload::ClL2NormalizationFloatWorkload(const L2NormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) - : FloatWorkload<L2NormalizationQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClL2NormalizationFloatWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - int axis = (m_Data.m_Parameters.m_DataLayout == DataLayout::NCHW) ? 2 : 0; - - m_Layer.configure(&input, &output, axis, m_Data.m_Parameters.m_Eps); -} - -void ClL2NormalizationFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClL2NormalizationFloatWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClL2NormalizationFloatWorkload.hpp b/src/backends/cl/workloads/ClL2NormalizationFloatWorkload.hpp deleted file mode 100644 index 26aea9fd88..0000000000 --- a/src/backends/cl/workloads/ClL2NormalizationFloatWorkload.hpp +++ /dev/null @@ -1,35 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLL2NormalizeLayer.h> - -namespace armnn -{ - -arm_compute::Status ClL2NormalizationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const L2NormalizationDescriptor& descriptor); - -class ClL2NormalizationFloatWorkload : public FloatWorkload<L2NormalizationQueueDescriptor> -{ -public: - ClL2NormalizationFloatWorkload(const L2NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - // Purposely not a CLL2Normalize function. See constructor. - mutable arm_compute::CLL2NormalizeLayer m_Layer; -}; - -} //namespace armnn - - - - diff --git a/src/backends/cl/workloads/ClLstmFloatWorkload.cpp b/src/backends/cl/workloads/ClLstmFloatWorkload.cpp deleted file mode 100644 index 2f3ba75275..0000000000 --- a/src/backends/cl/workloads/ClLstmFloatWorkload.cpp +++ /dev/null @@ -1,435 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClLstmFloatWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <arm_compute/runtime/CL/functions/CLLSTMLayer.h> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -ClLstmFloatWorkload::ClLstmFloatWorkload(const LstmQueueDescriptor &descriptor, const WorkloadInfo &info) - : FloatWorkload<LstmQueueDescriptor>(descriptor, info) -{ - arm_compute::LSTMParams<arm_compute::ICLTensor> lstm_param; - - // Basic parameters - m_InputToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo()); - - m_InputToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo()); - - m_InputToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo()); - - m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo()); - - m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo()); - - m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo()); - - m_ForgetGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo()); - - m_CellBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo()); - - m_OutputGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo()); - - // for future reference: check the AndroidNN API for the logic here - if (!m_Data.m_Parameters.m_CifgEnabled) - { - m_InputToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo()); - - m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo()); - - m_CellToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - if (m_Data.m_CellToInputWeights != nullptr) - { - BuildArmComputeTensor(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights->GetTensorInfo()); - } - - m_InputGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo()); - - lstm_param.set_cifg_params(m_InputToInputWeightsTensor.get(), - m_RecurrentToInputWeightsTensor.get(), - m_Data.m_CellToInputWeights != nullptr ? m_CellToInputWeightsTensor.get() : nullptr, - m_InputGateBiasTensor.get()); - } - - if (m_Data.m_Parameters.m_ProjectionEnabled) - { - m_ProjectionWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights->GetTensorInfo()); - - m_ProjectionBiasTensor = std::make_unique<arm_compute::CLTensor>(); - if (m_Data.m_ProjectionBias != nullptr) - { - BuildArmComputeTensor(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias->GetTensorInfo()); - } - - lstm_param.set_projection_params(m_ProjectionWeightsTensor.get(), - m_Data.m_ProjectionBias != nullptr ? m_ProjectionBiasTensor.get() : nullptr); - } - - if (m_Data.m_Parameters.m_PeepholeEnabled) - { - m_CellToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights->GetTensorInfo()); - - m_CellToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights->GetTensorInfo()); - - lstm_param.set_peephole_params(m_CellToForgetWeightsTensor.get(), m_CellToOutputWeightsTensor.get()); - } - - if (m_Data.m_Parameters.m_LayerNormEnabled) - { - m_InputLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - m_ForgetLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - m_CellLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - m_OutputLayerNormWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - - if (!m_Data.m_Parameters.m_CifgEnabled) - { - BuildArmComputeTensor(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights->GetTensorInfo()); - } - BuildArmComputeTensor(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights->GetTensorInfo()); - BuildArmComputeTensor(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights->GetTensorInfo()); - BuildArmComputeTensor(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights->GetTensorInfo()); - - lstm_param.set_layer_normalization_params(m_Data.m_Parameters.m_CifgEnabled ? nullptr : - m_InputLayerNormWeightsTensor.get(), - m_ForgetLayerNormWeightsTensor.get(), - m_CellLayerNormWeightsTensor.get(), - m_OutputLayerNormWeightsTensor.get()); - } - - const arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - const arm_compute::ICLTensor& output_state_in = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - const arm_compute::ICLTensor& cell_state_in = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor(); - - arm_compute::ICLTensor& output_state_out = static_cast<IClTensorHandle*>(m_Data.m_Outputs[1])->GetTensor(); - arm_compute::ICLTensor& cell_state_out = static_cast<IClTensorHandle*>(m_Data.m_Outputs[2])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[3])->GetTensor(); - - // Get the batch_size and the num_units from the cellStateIn dimensions - const TensorInfo& inputTensorInfo = info.m_InputTensorInfos[2]; - const unsigned int batch_size = boost::numeric_cast<unsigned int>(inputTensorInfo.GetShape()[0]); - const unsigned int num_units = boost::numeric_cast<unsigned int>(inputTensorInfo.GetShape()[1]); - - m_ScratchBuffer = std::make_unique<arm_compute::CLTensor>(); - if (m_Data.m_Parameters.m_CifgEnabled) - { - // 2D tensor with dimensions [num_units * 3, batch_size] with CIFG - armnn::TensorInfo scratchBuffer1({ batch_size, num_units * 3 }, DataType::Float32); - BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer1); - } - else - { - // scratch_buffer [num_units * 4, batch_size] without CIFG - armnn::TensorInfo scratchBuffer2({ batch_size, num_units * 4 }, DataType::Float32); - BuildArmComputeTensor(*m_ScratchBuffer, scratchBuffer2); - } - - float cell_threshold = m_Data.m_Parameters.m_ClippingThresCell; - float projection_threshold = m_Data.m_Parameters.m_ClippingThresProj; - - // for preparing the object for the class ActivationLayerInfo, we need to consider 5 situations - arm_compute::ActivationLayerInfo activationLayerInfo; - if (m_Data.m_Parameters.m_ActivationFunc == 0) - { - // no activation, do nothing - } - else if (m_Data.m_Parameters.m_ActivationFunc == 1) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::RELU); - } - else if (m_Data.m_Parameters.m_ActivationFunc == 3) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.0); - } - else if (m_Data.m_Parameters.m_ActivationFunc == 4) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::TANH, 1.0, 1.0); - } - else if (m_Data.m_Parameters.m_ActivationFunc == 6) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::LOGISTIC); - } - else - { - throw armnn::Exception("Wrong Type of Activation Function!"); - } - - m_LstmLayer.configure(&input, m_InputToForgetWeightsTensor.get(), m_InputToCellWeightsTensor.get(), - m_InputToOutputWeightsTensor.get(), m_RecurrentToForgetWeightsTensor.get(), - m_RecurrentToCellWeightsTensor.get(), m_RecurrentToOutputWeightsTensor.get(), - m_ForgetGateBiasTensor.get(), m_CellBiasTensor.get(), m_OutputGateBiasTensor.get(), - &output_state_in, &cell_state_in, m_ScratchBuffer.get(), &output_state_out, - &cell_state_out, &output, lstm_param, activationLayerInfo, - cell_threshold, projection_threshold); - - armcomputetensorutils::InitialiseArmComputeTensorEmpty(*m_ScratchBuffer); - - InitializeArmComputeClTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights); - InitializeArmComputeClTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights); - InitializeArmComputeClTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights); - InitializeArmComputeClTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights); - InitializeArmComputeClTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights); - InitializeArmComputeClTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights); - InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias); - InitializeArmComputeClTensorData(*m_CellBiasTensor, m_Data.m_CellBias); - InitializeArmComputeClTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias); - - if (!m_Data.m_Parameters.m_CifgEnabled) - { - InitializeArmComputeClTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights); - InitializeArmComputeClTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights); - if (m_Data.m_CellToInputWeights != nullptr) - { - InitializeArmComputeClTensorData(*m_CellToInputWeightsTensor, m_Data.m_CellToInputWeights); - } - InitializeArmComputeClTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias); - } - - if (m_Data.m_Parameters.m_ProjectionEnabled) - { - InitializeArmComputeClTensorData(*m_ProjectionWeightsTensor, m_Data.m_ProjectionWeights); - if (m_Data.m_ProjectionBias != nullptr) - { - InitializeArmComputeClTensorData(*m_ProjectionBiasTensor, m_Data.m_ProjectionBias); - } - } - - if (m_Data.m_Parameters.m_PeepholeEnabled) - { - InitializeArmComputeClTensorData(*m_CellToForgetWeightsTensor, m_Data.m_CellToForgetWeights); - InitializeArmComputeClTensorData(*m_CellToOutputWeightsTensor, m_Data.m_CellToOutputWeights); - } - - if (m_Data.m_Parameters.m_LayerNormEnabled) - { - if (!m_Data.m_Parameters.m_CifgEnabled) - { - InitializeArmComputeClTensorData(*m_InputLayerNormWeightsTensor, m_Data.m_InputLayerNormWeights); - } - - InitializeArmComputeClTensorData(*m_ForgetLayerNormWeightsTensor, m_Data.m_ForgetLayerNormWeights); - InitializeArmComputeClTensorData(*m_CellLayerNormWeightsTensor, m_Data.m_CellLayerNormWeights); - InitializeArmComputeClTensorData(*m_OutputLayerNormWeightsTensor, m_Data.m_OutputLayerNormWeights); - } - - // Force Compute Library to perform the necessary copying and reshaping, after which - // delete all the input tensors that will no longer be needed - m_LstmLayer.prepare(); - FreeUnusedTensors(); -} - -void ClLstmFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClLstmFloatWorkload_Execute"); - RunClFunction(m_LstmLayer, CHECK_LOCATION()); -} - -arm_compute::Status ClLstmFloatWorkloadValidate(const TensorInfo& input, const TensorInfo& outputStateIn, - const TensorInfo& cellStateIn, const TensorInfo& scratchBuffer, - const TensorInfo& outputStateOut, const TensorInfo& cellStateOut, - const TensorInfo& output, const LstmDescriptor& descriptor, - const LstmInputParamsInfo& paramsInfo) -{ - arm_compute::LSTMParams<arm_compute::ITensorInfo> lstm_params_info; - - // The inputs and the outputs - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputStateInInfo = BuildArmComputeTensorInfo(outputStateIn); - const arm_compute::TensorInfo aclCellStateInInfo = BuildArmComputeTensorInfo(cellStateIn); - const arm_compute::TensorInfo aclScratchBufferInfo = BuildArmComputeTensorInfo(scratchBuffer); - const arm_compute::TensorInfo aclOutputStateOutInfo = BuildArmComputeTensorInfo(outputStateOut); - const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - // Basic parameters - const arm_compute::TensorInfo aclInputToForgetWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights()); - const arm_compute::TensorInfo aclInputToCellWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights()); - const arm_compute::TensorInfo aclInputToOutputWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights()); - const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights()); - const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights()); - const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights()); - const arm_compute::TensorInfo aclForgetGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias()); - const arm_compute::TensorInfo aclCellBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellBias()); - const arm_compute::TensorInfo aclOutputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias()); - - arm_compute::TensorInfo aclInputToInputWeightsInfo; - arm_compute::TensorInfo aclRecurrentToInputWeightsInfo; - arm_compute::TensorInfo aclCellToInputWeightsInfo; - arm_compute::TensorInfo aclInputGateBiasInfo; - arm_compute::TensorInfo aclProjectionWeightsInfo; - arm_compute::TensorInfo aclProjectionBiasInfo; - arm_compute::TensorInfo aclCellToForgetWeightsInfo; - arm_compute::TensorInfo aclCellToOutputWeightsInfo; - arm_compute::TensorInfo aclInputLayerNormWeightsInfo; - arm_compute::TensorInfo aclForgetLayerNormWeightsInfo; - arm_compute::TensorInfo aclCellLayerNormWeightsInfo; - arm_compute::TensorInfo aclOutputLayerNormWeightsInfo; - - if (!descriptor.m_CifgEnabled) - { - aclInputToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights()); - aclRecurrentToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights()); - - if (paramsInfo.m_CellToInputWeights != nullptr) - { - aclCellToInputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToInputWeights()); - } - aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias()); - lstm_params_info.set_cifg_params(&aclInputToInputWeightsInfo, &aclRecurrentToInputWeightsInfo, - paramsInfo.m_CellToInputWeights != nullptr ? - &aclCellToInputWeightsInfo: nullptr, - &aclInputGateBiasInfo); - } - - if (descriptor.m_ProjectionEnabled) - { - aclProjectionWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetProjectionWeights()); - - if (paramsInfo.m_ProjectionBias != nullptr) - { - aclProjectionBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias()); - } - lstm_params_info.set_projection_params(&aclProjectionWeightsInfo, - paramsInfo.m_ProjectionBias != nullptr ? - &aclProjectionBiasInfo: nullptr); - } - - if (descriptor.m_PeepholeEnabled) - { - aclCellToForgetWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToForgetWeights()); - aclCellToOutputWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellToOutputWeights()); - lstm_params_info.set_peephole_params(&aclCellToForgetWeightsInfo, &aclCellToOutputWeightsInfo); - } - - float cell_threshold = descriptor.m_ClippingThresCell; - float projection_threshold = descriptor.m_ClippingThresProj; - - // for preparing the object for the class ActivationLayerInfo, we need to consider 5 situations - arm_compute::ActivationLayerInfo activationLayerInfo; - if (descriptor.m_ActivationFunc == 0) - { - // no activation, do nothing - } - else if (descriptor.m_ActivationFunc == 1) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::RELU); - } - else if (descriptor.m_ActivationFunc == 3) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.0); - } - else if (descriptor.m_ActivationFunc == 4) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::TANH, 1.0, 1.0); - } - else if (descriptor.m_ActivationFunc == 6) - { - activationLayerInfo = arm_compute::ActivationLayerInfo( - arm_compute::ActivationLayerInfo::ActivationFunction::LOGISTIC); - } - else - { - throw armnn::Exception("Wrong Type of Activation Function!"); - } - - if (descriptor.m_LayerNormEnabled) - { - if (!descriptor.m_CifgEnabled) - { - aclInputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputLayerNormWeights()); - } - - aclForgetLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetLayerNormWeights()); - - aclCellLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellLayerNormWeights()); - - aclOutputLayerNormWeightsInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputLayerNormWeights()); - - lstm_params_info.set_layer_normalization_params(descriptor.m_CifgEnabled ? - nullptr : &aclInputLayerNormWeightsInfo, - &aclForgetLayerNormWeightsInfo, - &aclCellLayerNormWeightsInfo, - &aclOutputLayerNormWeightsInfo); - } - - return arm_compute::CLLSTMLayer::validate(&aclInputInfo, &aclInputToForgetWeightsInfo, - &aclInputToCellWeightsInfo, - &aclInputToOutputWeightsInfo, - &aclRecurrentToForgetWeightsInfo, - &aclRecurrentToCellWeightsInfo, - &aclRecurrentToOutputWeightsInfo, - &aclForgetGateBiasInfo, - &aclCellBiasInfo, - &aclOutputGateBiasInfo, - &aclOutputStateInInfo, &aclCellStateInInfo, - &aclScratchBufferInfo, &aclOutputStateOutInfo, - &aclCellStateOutInfo, &aclOutputInfo, - lstm_params_info, activationLayerInfo, - cell_threshold, projection_threshold); -} - -void ClLstmFloatWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_InputToInputWeightsTensor); - FreeTensorIfUnused(m_InputToForgetWeightsTensor); - FreeTensorIfUnused(m_InputToCellWeightsTensor); - FreeTensorIfUnused(m_InputToOutputWeightsTensor); - FreeTensorIfUnused(m_RecurrentToInputWeightsTensor); - FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor); - FreeTensorIfUnused(m_RecurrentToCellWeightsTensor); - FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor); - FreeTensorIfUnused(m_CellToInputWeightsTensor); - FreeTensorIfUnused(m_CellToForgetWeightsTensor); - FreeTensorIfUnused(m_CellToOutputWeightsTensor); - FreeTensorIfUnused(m_InputGateBiasTensor); - FreeTensorIfUnused(m_ForgetGateBiasTensor); - FreeTensorIfUnused(m_CellBiasTensor); - FreeTensorIfUnused(m_OutputGateBiasTensor); - FreeTensorIfUnused(m_ProjectionWeightsTensor); - FreeTensorIfUnused(m_ProjectionBiasTensor); - FreeTensorIfUnused(m_ScratchBuffer); - FreeTensorIfUnused(m_InputLayerNormWeightsTensor); - FreeTensorIfUnused(m_ForgetLayerNormWeightsTensor); - FreeTensorIfUnused(m_CellLayerNormWeightsTensor); - FreeTensorIfUnused(m_OutputLayerNormWeightsTensor); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClLstmFloatWorkload.hpp b/src/backends/cl/workloads/ClLstmFloatWorkload.hpp deleted file mode 100644 index b7cb408b40..0000000000 --- a/src/backends/cl/workloads/ClLstmFloatWorkload.hpp +++ /dev/null @@ -1,59 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Descriptors.hpp> -#include <armnn/LstmParams.hpp> -#include <backendsCommon/Workload.hpp> -#include <backendsCommon/WorkloadData.hpp> - -#include <arm_compute/runtime/CL/functions/CLLSTMLayer.h> - -namespace armnn -{ - -class ClLstmFloatWorkload : public FloatWorkload<LstmQueueDescriptor> -{ -public: - ClLstmFloatWorkload(const LstmQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLLSTMLayer m_LstmLayer; - - std::unique_ptr<arm_compute::CLTensor> m_InputToInputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputToForgetWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputToCellWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputToOutputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToInputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToForgetWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToCellWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToOutputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_CellToInputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_CellToForgetWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_CellToOutputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputGateBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_ForgetGateBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_CellBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_OutputGateBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_ProjectionWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_ProjectionBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputLayerNormWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_ForgetLayerNormWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_CellLayerNormWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_OutputLayerNormWeightsTensor; - - std::unique_ptr<arm_compute::CLTensor> m_ScratchBuffer; - - void FreeUnusedTensors(); -}; - -arm_compute::Status ClLstmFloatWorkloadValidate(const TensorInfo& input, const TensorInfo& outputStateIn, - const TensorInfo& cellStateIn, const TensorInfo& scratchBuffer, - const TensorInfo& outputStateOut, const TensorInfo& cellStateOut, - const TensorInfo& output, const LstmDescriptor &descriptor, - const LstmInputParamsInfo& paramsInfo); -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMaximumWorkload.cpp b/src/backends/cl/workloads/ClMaximumWorkload.cpp deleted file mode 100644 index cd3192d186..0000000000 --- a/src/backends/cl/workloads/ClMaximumWorkload.cpp +++ /dev/null @@ -1,58 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClMaximumWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClMaximumWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0); - const arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - const arm_compute::Status aclStatus = arm_compute::CLElementwiseMax::validate(&aclInput0Info, - &aclInput1Info, - &aclOutputInfo); - - return aclStatus; -} - -ClMaximumWorkload::ClMaximumWorkload(const MaximumQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<MaximumQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClMaximumWorkload", 2, 1); - - arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_MaximumLayer.configure(&input0, &input1, &output); -} - -void ClMaximumWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClMaximumWorkload_Execute"); - RunClFunction(m_MaximumLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMaximumWorkload.hpp b/src/backends/cl/workloads/ClMaximumWorkload.hpp deleted file mode 100644 index 18f67cddf6..0000000000 --- a/src/backends/cl/workloads/ClMaximumWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLElementwiseOperations.h> - -namespace armnn -{ - -arm_compute::Status ClMaximumWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output); - -class ClMaximumWorkload : public BaseWorkload<MaximumQueueDescriptor> -{ -public: - ClMaximumWorkload(const MaximumQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLElementwiseMax m_MaximumLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMeanWorkload.cpp b/src/backends/cl/workloads/ClMeanWorkload.cpp deleted file mode 100644 index 470b6a883d..0000000000 --- a/src/backends/cl/workloads/ClMeanWorkload.cpp +++ /dev/null @@ -1,52 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClMeanWorkload.hpp" - -#include <cl/ClTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClMeanValidate(const TensorInfo& input, - const TensorInfo& output, - const MeanDescriptor& desc) -{ - const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(aclInputInfo.num_dimensions(), - input.GetNumDimensions(), - desc.m_Axis); - - return arm_compute::CLReduceMean::validate(&aclInputInfo, coords, desc.m_KeepDims, &aclOutputInfo); -} - -ClMeanWorkload::ClMeanWorkload(const MeanQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<MeanQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClMeanWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::Coordinates coords = BuildArmComputeReductionCoordinates(input.info()->num_dimensions(), - info.m_InputTensorInfos[0].GetNumDimensions(), - m_Data.m_Parameters.m_Axis); - - m_Layer.configure(&input, coords, m_Data.m_Parameters.m_KeepDims, &output); -} - -void ClMeanWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClMeanWorkload_Execute"); - m_Layer.run(); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMeanWorkload.hpp b/src/backends/cl/workloads/ClMeanWorkload.hpp deleted file mode 100644 index 127c0548b1..0000000000 --- a/src/backends/cl/workloads/ClMeanWorkload.hpp +++ /dev/null @@ -1,31 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLReduceMean.h> - -namespace armnn -{ - -arm_compute::Status ClMeanValidate(const TensorInfo& input, - const TensorInfo& output, - const MeanDescriptor& desc); - -class ClMeanWorkload : public BaseWorkload<MeanQueueDescriptor> -{ -public: - ClMeanWorkload(const MeanQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - // Not using CLMeanStdDev, as 4D input tensor support for Mean has been added to a new function called CLReduceMean. - mutable arm_compute::CLReduceMean m_Layer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMinimumWorkload.cpp b/src/backends/cl/workloads/ClMinimumWorkload.cpp deleted file mode 100644 index 5f8dfdb8eb..0000000000 --- a/src/backends/cl/workloads/ClMinimumWorkload.cpp +++ /dev/null @@ -1,58 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClMinimumWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClMinimumWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0); - const arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - const arm_compute::Status aclStatus = arm_compute::CLElementwiseMin::validate(&aclInput0Info, - &aclInput1Info, - &aclOutputInfo); - - return aclStatus; -} - -ClMinimumWorkload::ClMinimumWorkload(const MinimumQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<MinimumQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClMinimumWorkload", 2, 1); - - arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_MinimumLayer.configure(&input0, &input1, &output); -} - -void ClMinimumWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClMinimumWorkload_Execute"); - RunClFunction(m_MinimumLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMinimumWorkload.hpp b/src/backends/cl/workloads/ClMinimumWorkload.hpp deleted file mode 100644 index 55d7eea9dd..0000000000 --- a/src/backends/cl/workloads/ClMinimumWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLElementwiseOperations.h> - -namespace armnn -{ - -arm_compute::Status ClMinimumWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output); - -class ClMinimumWorkload : public BaseWorkload<MinimumQueueDescriptor> -{ -public: - ClMinimumWorkload(const MinimumQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLElementwiseMin m_MinimumLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMultiplicationWorkload.cpp b/src/backends/cl/workloads/ClMultiplicationWorkload.cpp deleted file mode 100644 index f4e7b836ed..0000000000 --- a/src/backends/cl/workloads/ClMultiplicationWorkload.cpp +++ /dev/null @@ -1,58 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClMultiplicationWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ - -arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput1 = armcomputetensorutils::BuildArmComputeTensorInfo(input0); - const arm_compute::TensorInfo aclInput2 = armcomputetensorutils::BuildArmComputeTensorInfo(input1); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - // At the time of writing, configure() will fail if a rounding policy other than TO_ZERO is supplied to it, - // when providing a scale of 1.0 for F32 tensors, even though the provided rounding policy appears to be - // ignored for F32 tensors. - return arm_compute::CLPixelWiseMultiplication::validate(&aclInput1, - &aclInput2, - &aclOutput, - 1.0f, - arm_compute::ConvertPolicy::SATURATE, - arm_compute::RoundingPolicy::TO_ZERO); -} - - -ClMultiplicationWorkload::ClMultiplicationWorkload(const MultiplicationQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<MultiplicationQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClMultiplicationWorkload", 2, 1); - - arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - // Construct - m_PixelWiseMultiplication.configure(&input0, - &input1, - &output, - 1.0f, - arm_compute::ConvertPolicy::SATURATE, - arm_compute::RoundingPolicy::TO_NEAREST_EVEN); -} - -void ClMultiplicationWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClMultiplicationWorkload_Execute"); - RunClFunction(m_PixelWiseMultiplication, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClMultiplicationWorkload.hpp b/src/backends/cl/workloads/ClMultiplicationWorkload.hpp deleted file mode 100644 index 732bb16dcc..0000000000 --- a/src/backends/cl/workloads/ClMultiplicationWorkload.hpp +++ /dev/null @@ -1,34 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h> - -namespace armnn -{ - -arm_compute::Status ClMultiplicationWorkloadValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output); - -class ClMultiplicationWorkload : public BaseWorkload<MultiplicationQueueDescriptor> -{ -public: - ClMultiplicationWorkload(const MultiplicationQueueDescriptor& descriptor, const WorkloadInfo& info); - - using BaseWorkload<MultiplicationQueueDescriptor>::BaseWorkload; - void Execute() const override; - -private: - mutable arm_compute::CLPixelWiseMultiplication m_PixelWiseMultiplication; -}; - -} //namespace armnn - - - diff --git a/src/backends/cl/workloads/ClNormalizationFloatWorkload.cpp b/src/backends/cl/workloads/ClNormalizationFloatWorkload.cpp deleted file mode 100644 index 5f2fd7ab83..0000000000 --- a/src/backends/cl/workloads/ClNormalizationFloatWorkload.cpp +++ /dev/null @@ -1,55 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClNormalizationFloatWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include "ClWorkloadUtils.hpp" - -using namespace armnn::armcomputetensorutils; - -namespace armnn -{ - -arm_compute::Status ClNormalizationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const NormalizationDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - - arm_compute::NormalizationLayerInfo layerInfo = BuildArmComputeNormalizationLayerInfo(descriptor); - - return arm_compute::CLNormalizationLayer::validate(&aclInputInfo, &aclOutputInfo, layerInfo); -} - -ClNormalizationFloatWorkload::ClNormalizationFloatWorkload(const NormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) - : FloatWorkload<NormalizationQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClNormalizationFloatWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - arm_compute::NormalizationLayerInfo normalizationInfo = BuildArmComputeNormalizationLayerInfo(m_Data.m_Parameters); - - m_NormalizationLayer.configure(&input, &output, normalizationInfo); -}; - -void ClNormalizationFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClNormalizationFloatWorkload_Execute"); - RunClFunction(m_NormalizationLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClNormalizationFloatWorkload.hpp b/src/backends/cl/workloads/ClNormalizationFloatWorkload.hpp deleted file mode 100644 index a6d4f25e5e..0000000000 --- a/src/backends/cl/workloads/ClNormalizationFloatWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLNormalizationLayer.h> - -namespace armnn -{ - -arm_compute::Status ClNormalizationWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const NormalizationDescriptor& descriptor); - -class ClNormalizationFloatWorkload : public FloatWorkload<NormalizationQueueDescriptor> -{ -public: - ClNormalizationFloatWorkload(const NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLNormalizationLayer m_NormalizationLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClPadWorkload.cpp b/src/backends/cl/workloads/ClPadWorkload.cpp deleted file mode 100644 index 8a8c34a212..0000000000 --- a/src/backends/cl/workloads/ClPadWorkload.cpp +++ /dev/null @@ -1,67 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClPadWorkload.hpp" - -#include <cl/ClTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <arm_compute/core/Types.h> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -ClPadWorkload::ClPadWorkload(const PadQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<PadQueueDescriptor>(descriptor, info) -{ - this->m_Data.ValidateInputsOutputs("ClPadWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor(); - - std::vector<std::pair<unsigned int, unsigned int>> reversed_PadList(descriptor.m_Parameters.m_PadList.size()); - - std::reverse_copy(std::begin(descriptor.m_Parameters.m_PadList), - std::end(descriptor.m_Parameters.m_PadList), - std::begin(reversed_PadList)); - - arm_compute::PaddingList padList = static_cast<arm_compute::PaddingList>(reversed_PadList); - - arm_compute::PixelValue pixelValue = GetPixelValue(input, descriptor.m_Parameters.m_PadValue); - - m_Layer.configure(&input, &output, padList, pixelValue); -} - -void ClPadWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClPadWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -arm_compute::Status ClPadValidate(const TensorInfo& input, - const TensorInfo& output, - const PadDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - std::vector<std::pair<unsigned int, unsigned int>> reversed_PadList(descriptor.m_PadList.size()); - - std::reverse_copy(std::begin(descriptor.m_PadList), - std::end(descriptor.m_PadList), - std::begin(reversed_PadList)); - - arm_compute::PaddingList padList = static_cast<arm_compute::PaddingList>(reversed_PadList); - - const arm_compute::Status aclStatus = arm_compute::CLPadLayer::validate(&aclInputInfo, - &aclOutputInfo, - padList); - - return aclStatus; -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClPadWorkload.hpp b/src/backends/cl/workloads/ClPadWorkload.hpp deleted file mode 100644 index e87cbb71c2..0000000000 --- a/src/backends/cl/workloads/ClPadWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/WorkloadData.hpp> -#include <backendsCommon/Workload.hpp> -#include <arm_compute/runtime/CL/functions/CLPadLayer.h> - -namespace armnn { - -class ClPadWorkload : public BaseWorkload<PadQueueDescriptor> -{ -public: - ClPadWorkload(const PadQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable arm_compute::CLPadLayer m_Layer; -}; - -arm_compute::Status ClPadValidate(const TensorInfo& input, - const TensorInfo& output, - const PadDescriptor& descriptor); - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClPermuteWorkload.cpp b/src/backends/cl/workloads/ClPermuteWorkload.cpp deleted file mode 100644 index 41bce1d4fa..0000000000 --- a/src/backends/cl/workloads/ClPermuteWorkload.cpp +++ /dev/null @@ -1,51 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClPermuteWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <arm_compute/core/Error.h> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ - -arm_compute::Status ClPermuteWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const PermuteDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output); - const armnn::PermutationVector& mappings = descriptor.m_DimMappings; - - return arm_compute::CLPermute::validate(&aclInputInfo, &aclOutputInfo, - armcomputetensorutils::BuildArmComputePermutationVector(mappings)); -} - -ClPermuteWorkload::ClPermuteWorkload(const PermuteQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<PermuteQueueDescriptor>(descriptor, info) -{ - using armcomputetensorutils::BuildArmComputePermutationVector; - - m_Data.ValidateInputsOutputs(GetName(), 1, 1); - - const arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - const armnn::PermutationVector& mappings = m_Data.m_Parameters.m_DimMappings; - - // Run the layer. - m_PermuteFunction.configure(&input, &output, BuildArmComputePermutationVector(mappings)); -} - -void ClPermuteWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL( GetName() + "_Execute"); - RunClFunction(m_PermuteFunction, CHECK_LOCATION()); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClPermuteWorkload.hpp b/src/backends/cl/workloads/ClPermuteWorkload.hpp deleted file mode 100644 index 8b5f4c6147..0000000000 --- a/src/backends/cl/workloads/ClPermuteWorkload.hpp +++ /dev/null @@ -1,40 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> -#include <backendsCommon/WorkloadData.hpp> - -#include <armnn/TypesUtils.hpp> -#include <arm_compute/runtime/CL/functions/CLPermute.h> - -#include <string> - -namespace armnn -{ - -arm_compute::Status ClPermuteWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const PermuteDescriptor& descriptor); - -class ClPermuteWorkload : public BaseWorkload<PermuteQueueDescriptor> -{ -public: - static const std::string& GetName() - { - static const std::string name = std::string("ClPermuteWorkload"); - return name; - } - - ClPermuteWorkload(const PermuteQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - using BaseWorkload<PermuteQueueDescriptor>::m_Data; - mutable arm_compute::CLPermute m_PermuteFunction; -}; - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClPooling2dWorkload.cpp b/src/backends/cl/workloads/ClPooling2dWorkload.cpp deleted file mode 100644 index 683880439f..0000000000 --- a/src/backends/cl/workloads/ClPooling2dWorkload.cpp +++ /dev/null @@ -1,60 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClPooling2dWorkload.hpp" -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClPooling2dWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const Pooling2dDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - - arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(descriptor); - - return arm_compute::CLPoolingLayer::validate(&aclInputInfo, &aclOutputInfo, layerInfo); -} - -ClPooling2dWorkload::ClPooling2dWorkload( - const Pooling2dQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<Pooling2dQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClPooling2dWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - // flag to use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy - // enable fp_mixed_precision for the the FP16 cases that - // accumulation reaches a limit beyond which there is no more increment of the value - bool fpMixedPrecision = false; - - arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(m_Data.m_Parameters, fpMixedPrecision); - - // Run the layer. - m_PoolingLayer.configure(&input, &output, layerInfo); -} - -void ClPooling2dWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClPooling2dWorkload_Execute"); - RunClFunction(m_PoolingLayer, CHECK_LOCATION()); -} - -} diff --git a/src/backends/cl/workloads/ClPooling2dWorkload.hpp b/src/backends/cl/workloads/ClPooling2dWorkload.hpp deleted file mode 100644 index ce67db2a13..0000000000 --- a/src/backends/cl/workloads/ClPooling2dWorkload.hpp +++ /dev/null @@ -1,33 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLPoolingLayer.h> - -namespace armnn -{ - -arm_compute::Status ClPooling2dWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const Pooling2dDescriptor& descriptor); - -class ClPooling2dWorkload : public BaseWorkload<Pooling2dQueueDescriptor> -{ -public: - using BaseWorkload<Pooling2dQueueDescriptor>::m_Data; - - ClPooling2dWorkload(const Pooling2dQueueDescriptor& descriptor, - const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable arm_compute::CLPoolingLayer m_PoolingLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClPreluWorkload.cpp b/src/backends/cl/workloads/ClPreluWorkload.cpp deleted file mode 100644 index 1813105c21..0000000000 --- a/src/backends/cl/workloads/ClPreluWorkload.cpp +++ /dev/null @@ -1,48 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClPreluWorkload.hpp" -#include "ClWorkloadUtils.hpp" -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeUtils.hpp> -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> - -namespace armnn -{ - -arm_compute::Status ClPreluWorkloadValidate(const TensorInfo& input, - const TensorInfo& alpha, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclAlpha = armcomputetensorutils::BuildArmComputeTensorInfo(alpha); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - return arm_compute::CLPReluLayer::validate(&aclInput, - &aclAlpha, - &aclOutput); -} - -ClPreluWorkload::ClPreluWorkload(const PreluQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<PreluQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClPreluWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& alpha = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_PreluLayer.configure(&input, &alpha, &output); -} - -void ClPreluWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClPreluWorkload_Execute"); - RunClFunction(m_PreluLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClPreluWorkload.hpp b/src/backends/cl/workloads/ClPreluWorkload.hpp deleted file mode 100644 index 9061416431..0000000000 --- a/src/backends/cl/workloads/ClPreluWorkload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLPReluLayer.h> - -namespace armnn -{ -arm_compute::Status ClPreluWorkloadValidate(const TensorInfo& input, - const TensorInfo& alpha, - const TensorInfo& output); - -class ClPreluWorkload : public BaseWorkload<PreluQueueDescriptor> -{ -public: - ClPreluWorkload(const PreluQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLPReluLayer m_PreluLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClQuantizeWorkload.cpp b/src/backends/cl/workloads/ClQuantizeWorkload.cpp deleted file mode 100644 index 263065a5a4..0000000000 --- a/src/backends/cl/workloads/ClQuantizeWorkload.cpp +++ /dev/null @@ -1,49 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClQuantizeWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClQuantizeWorkloadValidate(const TensorInfo& input, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - return arm_compute::CLQuantizationLayer::validate(&aclInputInfo, - &aclOutputInfo); -} - -ClQuantizeWorkload::ClQuantizeWorkload(const QuantizeQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<QuantizeQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClQuantizeWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_Layer.configure(&input, &output); -} - -void ClQuantizeWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClQuantizeWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClQuantizeWorkload.hpp b/src/backends/cl/workloads/ClQuantizeWorkload.hpp deleted file mode 100644 index f4a7ec64dd..0000000000 --- a/src/backends/cl/workloads/ClQuantizeWorkload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLQuantizationLayer.h> - -namespace armnn -{ - -arm_compute::Status ClQuantizeWorkloadValidate(const TensorInfo& input, - const TensorInfo& output); - -class ClQuantizeWorkload : public BaseWorkload<QuantizeQueueDescriptor> -{ -public: - ClQuantizeWorkload(const QuantizeQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLQuantizationLayer m_Layer; -}; - -} //namespace armnn
\ No newline at end of file diff --git a/src/backends/cl/workloads/ClQuantizedLstmWorkload.cpp b/src/backends/cl/workloads/ClQuantizedLstmWorkload.cpp deleted file mode 100644 index 688ebf9184..0000000000 --- a/src/backends/cl/workloads/ClQuantizedLstmWorkload.cpp +++ /dev/null @@ -1,158 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClQuantizedLstmWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <cl/ClTensorHandle.hpp> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClQuantizedLstmWorkloadValidate(const TensorInfo& input, const TensorInfo& previousCellStateIn, - const TensorInfo& previousOutputIn, const TensorInfo& cellStateOut, - const TensorInfo& output, - const QuantizedLstmInputParamsInfo& paramsInfo) -{ - // Inputs - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclPreviousCellStateInInfo = BuildArmComputeTensorInfo(previousCellStateIn); - const arm_compute::TensorInfo aclPreviousOutputInInfo = BuildArmComputeTensorInfo(previousOutputIn); - - // Outputs - const arm_compute::TensorInfo aclCellStateOutInfo = BuildArmComputeTensorInfo(cellStateOut); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - // Basic parameters - const arm_compute::TensorInfo aclInputToInputWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetInputToInputWeights()); - const arm_compute::TensorInfo aclInputToForgetWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetInputToForgetWeights()); - const arm_compute::TensorInfo aclInputToCellWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetInputToCellWeights()); - const arm_compute::TensorInfo aclInputToOutputWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetInputToOutputWeights()); - const arm_compute::TensorInfo aclRecurrentToInputWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToInputWeights()); - const arm_compute::TensorInfo aclRecurrentToForgetWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToForgetWeights()); - const arm_compute::TensorInfo aclRecurrentToCellWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToCellWeights()); - const arm_compute::TensorInfo aclRecurrentToOutputWeightsInfo - = BuildArmComputeTensorInfo(paramsInfo.GetRecurrentToOutputWeights()); - const arm_compute::TensorInfo aclInputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetInputGateBias()); - const arm_compute::TensorInfo aclForgetGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetForgetGateBias()); - const arm_compute::TensorInfo aclCellBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetCellBias()); - const arm_compute::TensorInfo aclOutputGateBiasInfo = BuildArmComputeTensorInfo(paramsInfo.GetOutputGateBias()); - - return arm_compute::CLLSTMLayerQuantized::validate(&aclInputInfo, &aclInputToInputWeightsInfo, - &aclInputToForgetWeightsInfo, &aclInputToCellWeightsInfo, - &aclInputToOutputWeightsInfo, &aclRecurrentToInputWeightsInfo, - &aclRecurrentToForgetWeightsInfo, &aclRecurrentToCellWeightsInfo, - &aclRecurrentToOutputWeightsInfo, &aclInputGateBiasInfo, - &aclForgetGateBiasInfo, &aclCellBiasInfo, &aclOutputGateBiasInfo, - &aclPreviousCellStateInInfo, &aclPreviousOutputInInfo, - &aclCellStateOutInfo, &aclOutputInfo); -} - -ClQuantizedLstmWorkload::ClQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor &descriptor, - const WorkloadInfo &info): - BaseWorkload<QuantizedLstmQueueDescriptor>(descriptor, info) -{ - m_InputToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights->GetTensorInfo()); - - m_InputToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights->GetTensorInfo()); - - m_InputToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights->GetTensorInfo()); - - m_InputToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights->GetTensorInfo()); - - m_RecurrentToInputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights->GetTensorInfo()); - - m_RecurrentToForgetWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights->GetTensorInfo()); - - m_RecurrentToCellWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights->GetTensorInfo()); - - m_RecurrentToOutputWeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights->GetTensorInfo()); - - m_InputGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_InputGateBiasTensor, m_Data.m_InputGateBias->GetTensorInfo()); - - m_ForgetGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias->GetTensorInfo()); - - m_CellBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_CellBiasTensor, m_Data.m_CellBias->GetTensorInfo()); - - m_OutputGateBiasTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias->GetTensorInfo()); - - const arm_compute::ICLTensor& inputTensor = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& cellStateInTensor = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); - const arm_compute::ICLTensor& outputStateInTensor = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor(); - - arm_compute::ICLTensor& cellStateOutTensor = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - arm_compute::ICLTensor& outputStateOutTensor = static_cast<IClTensorHandle*>(m_Data.m_Outputs[1])->GetTensor(); - - m_QuantizedLstmLayer.configure(&inputTensor, m_InputToInputWeightsTensor.get(), m_InputToForgetWeightsTensor.get(), - m_InputToCellWeightsTensor.get(), m_InputToOutputWeightsTensor.get(), - m_RecurrentToInputWeightsTensor.get(), m_RecurrentToForgetWeightsTensor.get(), - m_RecurrentToCellWeightsTensor.get(), m_RecurrentToOutputWeightsTensor.get(), - m_InputGateBiasTensor.get(), m_ForgetGateBiasTensor.get(), m_CellBiasTensor.get(), - m_OutputGateBiasTensor.get(), &cellStateInTensor, &outputStateInTensor, - &cellStateOutTensor, &outputStateOutTensor); - - InitializeArmComputeClTensorData(*m_InputToInputWeightsTensor, m_Data.m_InputToInputWeights); - InitializeArmComputeClTensorData(*m_InputToForgetWeightsTensor, m_Data.m_InputToForgetWeights); - InitializeArmComputeClTensorData(*m_InputToCellWeightsTensor, m_Data.m_InputToCellWeights); - InitializeArmComputeClTensorData(*m_InputToOutputWeightsTensor, m_Data.m_InputToOutputWeights); - InitializeArmComputeClTensorData(*m_RecurrentToInputWeightsTensor, m_Data.m_RecurrentToInputWeights); - InitializeArmComputeClTensorData(*m_RecurrentToForgetWeightsTensor, m_Data.m_RecurrentToForgetWeights); - InitializeArmComputeClTensorData(*m_RecurrentToCellWeightsTensor, m_Data.m_RecurrentToCellWeights); - InitializeArmComputeClTensorData(*m_RecurrentToOutputWeightsTensor, m_Data.m_RecurrentToOutputWeights); - InitializeArmComputeClTensorData(*m_InputGateBiasTensor, m_Data.m_InputGateBias); - InitializeArmComputeClTensorData(*m_ForgetGateBiasTensor, m_Data.m_ForgetGateBias); - InitializeArmComputeClTensorData(*m_CellBiasTensor, m_Data.m_CellBias); - InitializeArmComputeClTensorData(*m_OutputGateBiasTensor, m_Data.m_OutputGateBias); - - m_QuantizedLstmLayer.prepare(); - FreeUnusedTensors(); -} - -void ClQuantizedLstmWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClQuantizedLstmWorkload_Execute"); - RunClFunction(m_QuantizedLstmLayer, CHECK_LOCATION()); -} - -void ClQuantizedLstmWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_InputToInputWeightsTensor); - FreeTensorIfUnused(m_InputToForgetWeightsTensor); - FreeTensorIfUnused(m_InputToCellWeightsTensor); - FreeTensorIfUnused(m_InputToOutputWeightsTensor); - FreeTensorIfUnused(m_RecurrentToInputWeightsTensor); - FreeTensorIfUnused(m_RecurrentToForgetWeightsTensor); - FreeTensorIfUnused(m_RecurrentToCellWeightsTensor); - FreeTensorIfUnused(m_RecurrentToOutputWeightsTensor); - FreeTensorIfUnused(m_InputGateBiasTensor); - FreeTensorIfUnused(m_ForgetGateBiasTensor); - FreeTensorIfUnused(m_CellBiasTensor); - FreeTensorIfUnused(m_OutputGateBiasTensor); -} - -} // namespace armnn
\ No newline at end of file diff --git a/src/backends/cl/workloads/ClQuantizedLstmWorkload.hpp b/src/backends/cl/workloads/ClQuantizedLstmWorkload.hpp deleted file mode 100644 index 580db490d6..0000000000 --- a/src/backends/cl/workloads/ClQuantizedLstmWorkload.hpp +++ /dev/null @@ -1,49 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/QuantizedLstmParams.hpp> -#include <backendsCommon/Workload.hpp> -#include <backendsCommon/WorkloadData.hpp> - -#include <arm_compute/runtime/CL/functions/CLLSTMLayerQuantized.h> - -namespace armnn -{ - -arm_compute::Status ClQuantizedLstmWorkloadValidate(const TensorInfo& input, const TensorInfo& previousCellStateIn, - const TensorInfo& previousOutputIn, const TensorInfo& cellStateOut, - const TensorInfo& output, - const QuantizedLstmInputParamsInfo& paramsInfo); - -class ClQuantizedLstmWorkload : public BaseWorkload<QuantizedLstmQueueDescriptor> -{ -public: - ClQuantizedLstmWorkload(const QuantizedLstmQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLLSTMLayerQuantized m_QuantizedLstmLayer; - - std::unique_ptr<arm_compute::CLTensor> m_InputToInputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputToForgetWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputToCellWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputToOutputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToInputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToForgetWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToCellWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_RecurrentToOutputWeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_InputGateBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_ForgetGateBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_CellBiasTensor; - std::unique_ptr<arm_compute::CLTensor> m_OutputGateBiasTensor; - - void FreeUnusedTensors(); -}; - -} //namespace armnn - - diff --git a/src/backends/cl/workloads/ClReshapeWorkload.cpp b/src/backends/cl/workloads/ClReshapeWorkload.cpp deleted file mode 100644 index d752290444..0000000000 --- a/src/backends/cl/workloads/ClReshapeWorkload.cpp +++ /dev/null @@ -1,41 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClReshapeWorkload.hpp" -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ - -arm_compute::Status ClReshapeWorkloadValidate(const TensorInfo& input, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - return arm_compute::CLReshapeLayer::validate(&aclInputInfo, &aclOutputInfo); -} - -ClReshapeWorkload::ClReshapeWorkload(const ReshapeQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<ReshapeQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClReshapeWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_Layer.configure(&input, &output); -} - -void ClReshapeWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClReshapeWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClReshapeWorkload.hpp b/src/backends/cl/workloads/ClReshapeWorkload.hpp deleted file mode 100644 index d836f1e43a..0000000000 --- a/src/backends/cl/workloads/ClReshapeWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLReshapeLayer.h> - -namespace armnn -{ - -arm_compute::Status ClReshapeWorkloadValidate(const TensorInfo& input, - const TensorInfo& output); - -class ClReshapeWorkload : public BaseWorkload<ReshapeQueueDescriptor> -{ -public: - ClReshapeWorkload(const ReshapeQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable arm_compute::CLReshapeLayer m_Layer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClResizeWorkload.cpp b/src/backends/cl/workloads/ClResizeWorkload.cpp deleted file mode 100644 index 05b212c3d4..0000000000 --- a/src/backends/cl/workloads/ClResizeWorkload.cpp +++ /dev/null @@ -1,77 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClResizeWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <cl/ClTensorHandle.hpp> - -using namespace armnn::armcomputetensorutils; - -namespace armnn -{ - -arm_compute::Status ClResizeWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const ResizeDescriptor& descriptor) -{ - arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(descriptor.m_DataLayout); - aclInputInfo.set_data_layout(aclDataLayout); - aclOutputInfo.set_data_layout(aclDataLayout); - - arm_compute::InterpolationPolicy aclInterpolationPolicy = - ConvertResizeMethodToAclInterpolationPolicy(descriptor.m_Method); - - return arm_compute::CLScale::validate(&aclInputInfo, - &aclOutputInfo, - aclInterpolationPolicy, - arm_compute::BorderMode::REPLICATE, - arm_compute::PixelValue(0.f), - arm_compute::SamplingPolicy::TOP_LEFT, - true, - descriptor.m_BilinearAlignCorners); -} - -ClResizeWorkload::ClResizeWorkload(const ResizeQueueDescriptor& descriptor, const WorkloadInfo& info) : - BaseWorkload<ResizeQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClResizeWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - arm_compute::InterpolationPolicy aclInterpolationPolicy = - ConvertResizeMethodToAclInterpolationPolicy(descriptor.m_Parameters.m_Method); - - m_ResizeLayer.configure(&input, - &output, - aclInterpolationPolicy, - arm_compute::BorderMode::REPLICATE, - arm_compute::PixelValue(0.f), - arm_compute::SamplingPolicy::TOP_LEFT, - true, - descriptor.m_Parameters.m_BilinearAlignCorners); -}; - -void ClResizeWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClResizeWorkload_Execute"); - RunClFunction(m_ResizeLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClResizeWorkload.hpp b/src/backends/cl/workloads/ClResizeWorkload.hpp deleted file mode 100644 index ab5b943457..0000000000 --- a/src/backends/cl/workloads/ClResizeWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLScale.h> - -namespace armnn -{ - -arm_compute::Status ClResizeWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const ResizeDescriptor& descriptor); - -class ClResizeWorkload : public BaseWorkload<ResizeQueueDescriptor> -{ -public: - ClResizeWorkload(const ResizeQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLScale m_ResizeLayer; -}; - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClRsqrtWorkload.cpp b/src/backends/cl/workloads/ClRsqrtWorkload.cpp deleted file mode 100644 index be687595f7..0000000000 --- a/src/backends/cl/workloads/ClRsqrtWorkload.cpp +++ /dev/null @@ -1,44 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClRsqrtWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <cl/ClTensorHandle.hpp> - -#include <boost/cast.hpp> - -namespace armnn -{ - -arm_compute::Status ClRsqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - return arm_compute::CLRsqrtLayer::validate(&aclInput, &aclOutput); -} - -ClRsqrtWorkload::ClRsqrtWorkload(const RsqrtQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<RsqrtQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClRsqrtWorkload", 1, 1); - - arm_compute::ICLTensor& input = boost::polymorphic_downcast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = boost::polymorphic_downcast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - m_RsqrtLayer.configure(&input, &output); -} - -void ClRsqrtWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClRsqrtWorkload_Execute"); - RunClFunction(m_RsqrtLayer, CHECK_LOCATION()); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClRsqrtWorkload.hpp b/src/backends/cl/workloads/ClRsqrtWorkload.hpp deleted file mode 100644 index 8fb6229160..0000000000 --- a/src/backends/cl/workloads/ClRsqrtWorkload.hpp +++ /dev/null @@ -1,28 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/core/Error.h> -#include <arm_compute/runtime/CL/functions/CLElementWiseUnaryLayer.h> - -namespace armnn -{ - -arm_compute::Status ClRsqrtWorkloadValidate(const TensorInfo& input, const TensorInfo& output); - -class ClRsqrtWorkload : public BaseWorkload<RsqrtQueueDescriptor> -{ -public: - ClRsqrtWorkload(const RsqrtQueueDescriptor& descriptor, const WorkloadInfo& info); - virtual void Execute() const override; - -private: - mutable arm_compute::CLRsqrtLayer m_RsqrtLayer; -}; - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClSliceWorkload.cpp b/src/backends/cl/workloads/ClSliceWorkload.cpp deleted file mode 100644 index fa99e7f54d..0000000000 --- a/src/backends/cl/workloads/ClSliceWorkload.cpp +++ /dev/null @@ -1,56 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSliceWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <cl/ClTensorHandle.hpp> - -#include <boost/cast.hpp> - -namespace armnn -{ - -arm_compute::Status ClSliceWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SliceDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInput = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutput = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - arm_compute::Coordinates starts; - arm_compute::Coordinates ends; - - std::tie(starts, ends) = SetClSliceData(descriptor.m_Begin, descriptor.m_Size); - - return arm_compute::CLSlice::validate(&aclInput, &aclOutput, starts, ends); -} - -ClSliceWorkload::ClSliceWorkload(const SliceQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<SliceQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClSliceWorkload", 1, 1); - - arm_compute::ICLTensor& input = boost::polymorphic_downcast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = boost::polymorphic_downcast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::Coordinates starts; - arm_compute::Coordinates ends; - - std::tie(starts, ends) = SetClSliceData(m_Data.m_Parameters.m_Begin, m_Data.m_Parameters.m_Size); - - m_SliceFunction.configure(&input, &output, starts, ends); -} - -void ClSliceWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSliceWorkload_Execute"); - RunClFunction(m_SliceFunction, CHECK_LOCATION()); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClSliceWorkload.hpp b/src/backends/cl/workloads/ClSliceWorkload.hpp deleted file mode 100644 index 3460b7788b..0000000000 --- a/src/backends/cl/workloads/ClSliceWorkload.hpp +++ /dev/null @@ -1,30 +0,0 @@ -// -// Copyright © 2019 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/core/Error.h> -#include <arm_compute/runtime/CL/functions/CLSlice.h> - -namespace armnn -{ - -arm_compute::Status ClSliceWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SliceDescriptor& descriptor); - -class ClSliceWorkload : public BaseWorkload<SliceQueueDescriptor> -{ -public: - ClSliceWorkload(const SliceQueueDescriptor& descriptor, const WorkloadInfo& info); - virtual void Execute() const override; - -private: - mutable arm_compute::CLSlice m_SliceFunction; -}; - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp deleted file mode 100644 index a355ba0c2d..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.cpp +++ /dev/null @@ -1,27 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSoftmaxBaseWorkload.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <aclCommon/ArmComputeUtils.hpp> - -#include <arm_compute/runtime/CL/functions/CLSoftmaxLayer.h> - -namespace armnn -{ - -arm_compute::Status ClSoftmaxWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SoftmaxDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(input); - const arm_compute::TensorInfo aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output); - - unsigned int aclAxis = ComputeSoftmaxAclAxis(descriptor, input); - return arm_compute::CLSoftmaxLayer::validate(&aclInputInfo, &aclOutputInfo, descriptor.m_Beta, aclAxis); -} - -} diff --git a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.hpp b/src/backends/cl/workloads/ClSoftmaxBaseWorkload.hpp deleted file mode 100644 index 8d73060162..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxBaseWorkload.hpp +++ /dev/null @@ -1,19 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Descriptors.hpp> -#include <armnn/Tensor.hpp> -#include <arm_compute/core/Error.h> - -namespace armnn -{ - -arm_compute::Status ClSoftmaxWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SoftmaxDescriptor& descriptor); - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp deleted file mode 100644 index adb4872b80..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.cpp +++ /dev/null @@ -1,36 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSoftmaxFloatWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -namespace armnn -{ - -ClSoftmaxFloatWorkload::ClSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) - : FloatWorkload<SoftmaxQueueDescriptor>(descriptor, info) - , m_SoftmaxLayer(memoryManager) -{ - m_Data.ValidateInputsOutputs("ClSoftmaxFloatWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]); - m_SoftmaxLayer.configure(&input, &output, m_Data.m_Parameters.m_Beta, aclAxis); -} - -void ClSoftmaxFloatWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSoftmaxFloatWorkload_Execute"); - RunClFunction(m_SoftmaxLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.hpp b/src/backends/cl/workloads/ClSoftmaxFloatWorkload.hpp deleted file mode 100644 index 7efdae858a..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxFloatWorkload.hpp +++ /dev/null @@ -1,30 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLSoftmaxLayer.h> -#include <arm_compute/runtime/MemoryManagerOnDemand.h> - -#include <memory> - -namespace armnn -{ - -class ClSoftmaxFloatWorkload : public FloatWorkload<SoftmaxQueueDescriptor> -{ -public: - ClSoftmaxFloatWorkload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); - void Execute() const override; - -private: - mutable arm_compute::CLSoftmaxLayer m_SoftmaxLayer; -}; - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp b/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp deleted file mode 100644 index f14ea11c82..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxUint8Workload.cpp +++ /dev/null @@ -1,46 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSoftmaxUint8Workload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -namespace armnn -{ - -ClSoftmaxUint8Workload::ClSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) - : Uint8Workload<SoftmaxQueueDescriptor>(descriptor, info) - , m_SoftmaxLayer(memoryManager) -{ - m_Data.ValidateInputsOutputs("ClSoftmaxUint8Workload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - const auto outputQuantization = output.info()->quantization_info(); - - if ((!outputQuantization.scale().empty() && outputQuantization.scale()[0] != (1.0f / 256.0f)) || - (!outputQuantization.offset().empty() && outputQuantization.offset()[0] != 0) || - outputQuantization.scale().empty() || outputQuantization.offset().empty()) - { - throw InvalidArgumentException( - "Invalid quantization for output. Only scale = 1.0f / 256.0f and offset = 0 supported"); - } - - unsigned int aclAxis = ComputeSoftmaxAclAxis(m_Data.m_Parameters, info.m_InputTensorInfos[0]); - m_SoftmaxLayer.configure(&input, &output, descriptor.m_Parameters.m_Beta, aclAxis); -} - -void ClSoftmaxUint8Workload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSoftmaxUint8Workload_Execute"); - RunClFunction(m_SoftmaxLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSoftmaxUint8Workload.hpp b/src/backends/cl/workloads/ClSoftmaxUint8Workload.hpp deleted file mode 100644 index f378b89a5d..0000000000 --- a/src/backends/cl/workloads/ClSoftmaxUint8Workload.hpp +++ /dev/null @@ -1,31 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLSoftmaxLayer.h> -#include <arm_compute/runtime/MemoryManagerOnDemand.h> - -#include <memory> - -namespace armnn -{ -// Softmax -class ClSoftmaxUint8Workload : public Uint8Workload<SoftmaxQueueDescriptor> -{ -public: - ClSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); - - void Execute() const override; -private: - - mutable arm_compute::CLSoftmaxLayer m_SoftmaxLayer; -}; - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClSpaceToBatchNdWorkload.cpp b/src/backends/cl/workloads/ClSpaceToBatchNdWorkload.cpp deleted file mode 100644 index 64da92c815..0000000000 --- a/src/backends/cl/workloads/ClSpaceToBatchNdWorkload.cpp +++ /dev/null @@ -1,85 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSpaceToBatchNdWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -#include <boost/polymorphic_pointer_cast.hpp> - -namespace armnn -{ -using namespace armcomputetensorutils; - -arm_compute::Status ClSpaceToBatchNdWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SpaceToBatchNdDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - - // ArmNN blockShape is [H, W] Cl asks for W, H - int32_t blockHeight = boost::numeric_cast<int32_t>(descriptor.m_BlockShape[0]); - int32_t blockWidth = boost::numeric_cast<int32_t>(descriptor.m_BlockShape[1]); - - arm_compute::Size2D paddingLeftTop = BuildArmComputeSize2D( - descriptor.m_PadList[1].first, descriptor.m_PadList[0].first); - arm_compute::Size2D paddingRightBottom = BuildArmComputeSize2D( - descriptor.m_PadList[1].second, descriptor.m_PadList[0].second); - - return arm_compute::CLSpaceToBatchLayer::validate(&aclInputInfo, - blockWidth, - blockHeight, - paddingLeftTop, - paddingRightBottom, - &aclOutputInfo); -} - -ClSpaceToBatchNdWorkload::ClSpaceToBatchNdWorkload( - const SpaceToBatchNdQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<SpaceToBatchNdQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClSpaceToBatchNdWorkload", 1, 1); - - arm_compute::ICLTensor& input = - boost::polymorphic_pointer_downcast<IClTensorHandle>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = - boost::polymorphic_pointer_downcast<IClTensorHandle>(m_Data.m_Outputs[0])->GetTensor(); - - // ArmNN blockShape is [H, W] Cl asks for W, H - int32_t blockHeight = boost::numeric_cast<int32_t>(m_Data.m_Parameters.m_BlockShape[0]); - int32_t blockWidth = boost::numeric_cast<int32_t>(m_Data.m_Parameters.m_BlockShape[1]); - - arm_compute::Size2D paddingLeftTop = BuildArmComputeSize2D( - m_Data.m_Parameters.m_PadList[1].first, m_Data.m_Parameters.m_PadList[0].first); - arm_compute::Size2D paddingRightBottom = BuildArmComputeSize2D( - m_Data.m_Parameters.m_PadList[1].second, m_Data.m_Parameters.m_PadList[0].second); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - m_SpaceToBatchLayer.configure(&input, - blockWidth, - blockHeight, - paddingLeftTop, - paddingRightBottom, - &output); -} - -void ClSpaceToBatchNdWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSpaceToBatchNdWorkload_Execute"); - RunClFunction(m_SpaceToBatchLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSpaceToBatchNdWorkload.hpp b/src/backends/cl/workloads/ClSpaceToBatchNdWorkload.hpp deleted file mode 100644 index 7500b5a3ac..0000000000 --- a/src/backends/cl/workloads/ClSpaceToBatchNdWorkload.hpp +++ /dev/null @@ -1,33 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Tensor.hpp> -#include <armnn/Descriptors.hpp> - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLSpaceToBatchLayer.h> - -namespace armnn -{ - -arm_compute::Status ClSpaceToBatchNdWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SpaceToBatchNdDescriptor& descriptor); - -class ClSpaceToBatchNdWorkload : public BaseWorkload<SpaceToBatchNdQueueDescriptor> -{ -public: - ClSpaceToBatchNdWorkload(const SpaceToBatchNdQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLSpaceToBatchLayer m_SpaceToBatchLayer; -}; - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClSpaceToDepthWorkload.cpp b/src/backends/cl/workloads/ClSpaceToDepthWorkload.cpp deleted file mode 100644 index d541e4ec52..0000000000 --- a/src/backends/cl/workloads/ClSpaceToDepthWorkload.cpp +++ /dev/null @@ -1,61 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSpaceToDepthWorkload.hpp" -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClTensorHandle.hpp> -#include <boost/polymorphic_pointer_cast.hpp> - -namespace armnn -{ -using namespace armcomputetensorutils; - -ClSpaceToDepthWorkload::ClSpaceToDepthWorkload(const SpaceToDepthQueueDescriptor& desc, - const WorkloadInfo& info) - : BaseWorkload<SpaceToDepthQueueDescriptor>(desc, info) -{ - m_Data.ValidateInputsOutputs("ClSpaceToDepthWorkload", 1, 1); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - input.info()->set_data_layout(aclDataLayout); - - int32_t blockSize = boost::numeric_cast<int32_t>(desc.m_Parameters.m_BlockSize); - - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - output.info()->set_data_layout(aclDataLayout); - - m_Layer.configure(&input, &output, blockSize); -} - -void ClSpaceToDepthWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSpaceToDepthWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -arm_compute::Status ClSpaceToDepthWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SpaceToDepthDescriptor& desc) -{ - DataLayout dataLayout = desc.m_DataLayout; - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, dataLayout); - - int32_t blockSize = boost::numeric_cast<int32_t>(desc.m_BlockSize); - - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, dataLayout); - - const arm_compute::Status aclStatus = arm_compute::CLSpaceToDepthLayer::validate(&aclInputInfo, - &aclOutputInfo, - blockSize); - return aclStatus; -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSpaceToDepthWorkload.hpp b/src/backends/cl/workloads/ClSpaceToDepthWorkload.hpp deleted file mode 100644 index 57ce5d4d05..0000000000 --- a/src/backends/cl/workloads/ClSpaceToDepthWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Descriptors.hpp> - -#include <backendsCommon/Workload.hpp> -#include <arm_compute/runtime/CL/functions/CLSpaceToDepthLayer.h> - -namespace armnn -{ -arm_compute::Status ClSpaceToDepthWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const SpaceToDepthDescriptor& desc); - -class ClSpaceToDepthWorkload : public BaseWorkload<SpaceToDepthQueueDescriptor> -{ -public: - ClSpaceToDepthWorkload(const SpaceToDepthQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLSpaceToDepthLayer m_Layer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSplitterWorkload.cpp b/src/backends/cl/workloads/ClSplitterWorkload.cpp deleted file mode 100644 index 296e0a3dde..0000000000 --- a/src/backends/cl/workloads/ClSplitterWorkload.cpp +++ /dev/null @@ -1,115 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSplitterWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <aclCommon/ArmComputeUtils.hpp> -#include <arm_compute/runtime/CL/functions/CLSplit.h> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClTensorHandle.hpp> - - -namespace armnn -{ - -using namespace armcomputetensorutils; - -namespace -{ - unsigned int CalcAclAxis(unsigned int numDimensions, unsigned int splitAxis) - { - return (numDimensions - splitAxis) - 1; - } - -} //namespace - -arm_compute::Status ClSplitterWorkloadValidate(const TensorInfo& input, - const std::vector<std::reference_wrapper<TensorInfo>>& outputs, - unsigned int splitAxis) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); - - size_t numOutputs = outputs.size(); - - std::vector<arm_compute::TensorInfo> aclOutputs; - aclOutputs.reserve(numOutputs); - - std::vector<arm_compute::ITensorInfo*> aclOutputPtr; - aclOutputPtr.reserve(numOutputs); - - for (size_t i = 0u; i < outputs.size(); ++i) - { - aclOutputs.emplace_back(BuildArmComputeTensorInfo(outputs[i])); - aclOutputPtr.emplace_back(&aclOutputs.back()); - } - - unsigned int aclAxis = CalcAclAxis(input.GetNumDimensions(), splitAxis); - return arm_compute::CLSplit::validate(&aclInputInfo, aclOutputPtr, aclAxis); -} - -ClSplitterWorkload::ClSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info) - : BaseWorkload<SplitterQueueDescriptor>(descriptor, info) -{ - bool allOutputsAreSubtensors = true; - - // Check that all outputs are sub-tensors - for (auto output : m_Data.m_Outputs) - { - if (output && !output->GetParent()) - { - // Non sub-tensor input found so we need to execute the split function - allOutputsAreSubtensors = false; - break; - } - } - - if (allOutputsAreSubtensors) - { - // Can skip configuring the split function since it's not executed - return; - } - - arm_compute::ICLTensor& input = boost::polymorphic_pointer_downcast<IClTensorHandle>( - m_Data.m_Inputs[0])->GetTensor(); - - std::vector<arm_compute::ICLTensor *> aclOutputs; - for (auto output : m_Data.m_Outputs) - { - arm_compute::ICLTensor& aclOutput = boost::polymorphic_pointer_downcast<IClTensorHandle>(output)->GetTensor(); - aclOutputs.emplace_back(&aclOutput); - } - - // Create the layer function - - // Configure input and output tensors - std::set<unsigned int> splitAxis = ComputeSplitAxis(descriptor.m_Parameters, m_Data.m_Inputs[0]->GetShape()); - if (splitAxis.size() != 1) - { - throw InvalidArgumentException("Cannot derive split axis from SplitterDescriptor"); - } - - unsigned int aclAxis = CalcAclAxis(descriptor.m_Parameters.GetNumDimensions(), *splitAxis.begin()); - auto layer = std::make_unique<arm_compute::CLSplit>(); - layer->configure(&input, aclOutputs, aclAxis); - - // Prepare - layer->prepare(); - - m_Layer = std::move(layer); -} - -void ClSplitterWorkload::Execute() const -{ - if (m_Layer) - { - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSplitterWorkload_Execute"); - m_Layer->run(); - } -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSplitterWorkload.hpp b/src/backends/cl/workloads/ClSplitterWorkload.hpp deleted file mode 100644 index 82211f5baf..0000000000 --- a/src/backends/cl/workloads/ClSplitterWorkload.hpp +++ /dev/null @@ -1,33 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/core/Error.h> -#include <arm_compute/runtime/IFunction.h> - -#include <functional> - -namespace armnn -{ - -arm_compute::Status ClSplitterWorkloadValidate(const TensorInfo& input, - const std::vector<std::reference_wrapper<TensorInfo>>& outputs, - unsigned int splitAxis); - -class ClSplitterWorkload : public BaseWorkload<SplitterQueueDescriptor> -{ -public: - ClSplitterWorkload(const SplitterQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - std::unique_ptr<arm_compute::IFunction> m_Layer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClStackWorkload.cpp b/src/backends/cl/workloads/ClStackWorkload.cpp deleted file mode 100644 index 3ba698ec4d..0000000000 --- a/src/backends/cl/workloads/ClStackWorkload.cpp +++ /dev/null @@ -1,74 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// -#include "ClStackWorkload.hpp" -#include "ClWorkloadUtils.hpp" -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -#include <arm_compute/core/Types.h> - -#include <boost/numeric/conversion/cast.hpp> -#include <boost/polymorphic_pointer_cast.hpp> - -namespace armnn -{ -using namespace armcomputetensorutils; - -namespace -{ -int CalcAxis(const unsigned int axis, const unsigned int inputDimensions) -{ - const int intAxis = boost::numeric_cast<int>(axis); - return boost::numeric_cast<int>(inputDimensions) - intAxis; -} -} //namespace - -arm_compute::Status ClStackWorkloadValidate(const std::vector<const TensorInfo*>& inputs, - const TensorInfo& output, - const StackDescriptor& descriptor) -{ - std::vector<arm_compute::ITensorInfo*> aclInputPtrs; - arm_compute::TensorInfo aclInputInfo; - for (const TensorInfo* input : inputs) - { - aclInputInfo = BuildArmComputeTensorInfo(*input); - aclInputPtrs.emplace_back(&aclInputInfo); - } - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - int aclAxis = CalcAxis(descriptor.m_Axis, descriptor.m_InputShape.GetNumDimensions()); - - return arm_compute::CLStackLayer::validate(aclInputPtrs, aclAxis, &aclOutputInfo); -} - -ClStackWorkload::ClStackWorkload(const StackQueueDescriptor& descriptor, const WorkloadInfo& info) -: BaseWorkload<StackQueueDescriptor>(descriptor, info) -{ - std::vector<arm_compute::ICLTensor*> aclInputs; - for (auto input : m_Data.m_Inputs) - { - arm_compute::ICLTensor& aclInput = boost::polymorphic_pointer_downcast<IClTensorHandle>(input)->GetTensor(); - aclInputs.emplace_back(&aclInput); - } - arm_compute::ICLTensor& output = boost::polymorphic_pointer_downcast<IClTensorHandle>( - m_Data.m_Outputs[0])->GetTensor(); - - m_Layer.reset(new arm_compute::CLStackLayer()); - int aclAxis = CalcAxis(descriptor.m_Parameters.m_Axis, descriptor.m_Parameters.m_InputShape.GetNumDimensions()); - m_Layer->configure(aclInputs, aclAxis, &output); -} - -void ClStackWorkload::Execute() const -{ - if (m_Layer) - { - ARMNN_SCOPED_PROFILING_EVENT_CL("ClStackWorkload_Execute"); - m_Layer->run(); - } -} - -} //namespace armnn
\ No newline at end of file diff --git a/src/backends/cl/workloads/ClStackWorkload.hpp b/src/backends/cl/workloads/ClStackWorkload.hpp deleted file mode 100644 index f27d6cdad0..0000000000 --- a/src/backends/cl/workloads/ClStackWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLStackLayer.h> - -namespace armnn -{ -arm_compute::Status ClStackWorkloadValidate(const std::vector<const TensorInfo*>& inputs, - const TensorInfo& output, - const StackDescriptor& descriptor); - -class ClStackWorkload : public BaseWorkload<StackQueueDescriptor> -{ -public: - ClStackWorkload(const StackQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable std::unique_ptr<arm_compute::CLStackLayer> m_Layer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClStridedSliceWorkload.cpp b/src/backends/cl/workloads/ClStridedSliceWorkload.cpp deleted file mode 100644 index 6b0a34d90e..0000000000 --- a/src/backends/cl/workloads/ClStridedSliceWorkload.cpp +++ /dev/null @@ -1,96 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClStridedSliceWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> -#include <backendsCommon/WorkloadUtils.hpp> - -#include <boost/numeric/conversion/cast.hpp> -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClStridedSliceWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const StridedSliceDescriptor& descriptor) -{ - const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - - arm_compute::Coordinates starts; - arm_compute::Coordinates ends; - arm_compute::Coordinates strides; - - std::tie(starts, ends, strides) = SetClStridedSliceData(descriptor.m_Begin, descriptor.m_End, descriptor.m_Stride); - - auto numDimensions = boost::numeric_cast<int>(input.GetNumDimensions()); - int32_t begin_mask = ConvertMaskToACLFormat(descriptor.m_BeginMask, numDimensions); - int32_t end_mask = ConvertMaskToACLFormat(descriptor.m_EndMask, numDimensions); - int32_t shrink_axis_mask = ConvertMaskToACLFormat(descriptor.m_ShrinkAxisMask, numDimensions); - - return arm_compute::CLStridedSlice::validate(&aclInputInfo, - &aclOutputInfo, - starts, - ends, - strides, - begin_mask, - end_mask, - shrink_axis_mask); -} - -ClStridedSliceWorkload::ClStridedSliceWorkload(const StridedSliceQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<StridedSliceQueueDescriptor>(descriptor, info) -{ - m_Data.ValidateInputsOutputs("ClStridedSliceWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::Coordinates starts; - arm_compute::Coordinates ends; - arm_compute::Coordinates strides; - - std::tie(starts, ends, strides) = SetClStridedSliceData(m_Data.m_Parameters.m_Begin, - m_Data.m_Parameters.m_End, - m_Data.m_Parameters.m_Stride); - - auto numDimensions = boost::numeric_cast<int>(info.m_InputTensorInfos[0].GetNumDimensions()); - int32_t begin_mask = ConvertMaskToACLFormat(m_Data.m_Parameters.m_BeginMask, numDimensions); - int32_t end_mask = ConvertMaskToACLFormat(m_Data.m_Parameters.m_EndMask, numDimensions); - int32_t shrink_axis_mask = ConvertMaskToACLFormat(m_Data.m_Parameters.m_ShrinkAxisMask, numDimensions); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - m_StridedSliceLayer.configure(&input, - &output, - starts, - ends, - strides, - begin_mask, - end_mask, - shrink_axis_mask); -} - -void ClStridedSliceWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClStridedSliceWorkload_Execute"); - RunClFunction(m_StridedSliceLayer, CHECK_LOCATION()); -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClStridedSliceWorkload.hpp b/src/backends/cl/workloads/ClStridedSliceWorkload.hpp deleted file mode 100644 index bce3fe13eb..0000000000 --- a/src/backends/cl/workloads/ClStridedSliceWorkload.hpp +++ /dev/null @@ -1,32 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Tensor.hpp> -#include <armnn/Descriptors.hpp> - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLStridedSlice.h> - -namespace armnn -{ - -arm_compute::Status ClStridedSliceWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const StridedSliceDescriptor& descriptor); - -class ClStridedSliceWorkload : public BaseWorkload<StridedSliceQueueDescriptor> -{ -public: - ClStridedSliceWorkload(const StridedSliceQueueDescriptor& descriptor, const WorkloadInfo& info); - void Execute() const override; - -private: - mutable arm_compute::CLStridedSlice m_StridedSliceLayer; -}; - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSubtractionWorkload.cpp b/src/backends/cl/workloads/ClSubtractionWorkload.cpp deleted file mode 100644 index 38154eb4d7..0000000000 --- a/src/backends/cl/workloads/ClSubtractionWorkload.cpp +++ /dev/null @@ -1,54 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClSubtractionWorkload.hpp" - -#include <cl/ClTensorHandle.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -static constexpr arm_compute::ConvertPolicy g_AclConvertPolicy = arm_compute::ConvertPolicy::SATURATE; - -ClSubtractionWorkload::ClSubtractionWorkload(const SubtractionQueueDescriptor& descriptor, - const WorkloadInfo& info) - : BaseWorkload<SubtractionQueueDescriptor>(descriptor, info) -{ - this->m_Data.ValidateInputsOutputs("ClSubtractionWorkload", 2, 1); - - arm_compute::ICLTensor& input0 = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& input1 = static_cast<IClTensorHandle*>(this->m_Data.m_Inputs[1])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(this->m_Data.m_Outputs[0])->GetTensor(); - m_Layer.configure(&input0, &input1, &output, g_AclConvertPolicy); -} - -void ClSubtractionWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClSubtractionWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -arm_compute::Status ClSubtractionValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output) -{ - const arm_compute::TensorInfo aclInput0Info = BuildArmComputeTensorInfo(input0); - const arm_compute::TensorInfo aclInput1Info = BuildArmComputeTensorInfo(input1); - const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); - - const arm_compute::Status aclStatus = arm_compute::CLArithmeticSubtraction::validate(&aclInput0Info, - &aclInput1Info, - &aclOutputInfo, - g_AclConvertPolicy); - - return aclStatus; -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClSubtractionWorkload.hpp b/src/backends/cl/workloads/ClSubtractionWorkload.hpp deleted file mode 100644 index da6d17c6ac..0000000000 --- a/src/backends/cl/workloads/ClSubtractionWorkload.hpp +++ /dev/null @@ -1,29 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLElementwiseOperations.h> - -namespace armnn -{ - -class ClSubtractionWorkload : public BaseWorkload<SubtractionQueueDescriptor> -{ -public: - ClSubtractionWorkload(const SubtractionQueueDescriptor& descriptor, const WorkloadInfo& info); - - void Execute() const override; - -private: - mutable arm_compute::CLArithmeticSubtraction m_Layer; -}; - -arm_compute::Status ClSubtractionValidate(const TensorInfo& input0, - const TensorInfo& input1, - const TensorInfo& output); -} //namespace armnn diff --git a/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.cpp b/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.cpp deleted file mode 100644 index 7c0736645b..0000000000 --- a/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.cpp +++ /dev/null @@ -1,110 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClTransposeConvolution2dWorkload.hpp" - -#include "ClWorkloadUtils.hpp" - -#include <cl/ClLayerSupport.hpp> -#include <cl/ClTensorHandle.hpp> -#include <cl/ClLayerSupport.hpp> - -#include <aclCommon/ArmComputeUtils.hpp> -#include <aclCommon/ArmComputeTensorUtils.hpp> - -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h> - -namespace armnn -{ - -using namespace armcomputetensorutils; - -arm_compute::Status ClTransposeConvolution2dWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const TransposeConvolution2dDescriptor& descriptor, - const TensorInfo& weights, - const Optional<TensorInfo>& biases) -{ - arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); - arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); - arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); - - arm_compute::TensorInfo aclBiasesInfo; - arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; - - if (descriptor.m_BiasEnabled) - { - BOOST_ASSERT(biases.has_value()); - - aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); - optionalAclBiasesInfo = &aclBiasesInfo; - } - - arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(descriptor); - - return arm_compute::CLDeconvolutionLayer::validate(&aclInputInfo, - &aclWeightsInfo, - optionalAclBiasesInfo, - &aclOutputInfo, - padStrideInfo); -} - -ClTransposeConvolution2dWorkload::ClTransposeConvolution2dWorkload( - const TransposeConvolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) : - BaseWorkload<TransposeConvolution2dQueueDescriptor>(descriptor, info), - m_Layer(memoryManager) -{ - const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo(); - - m_WeightsTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_WeightsTensor, weightInfo, m_Data.m_Parameters.m_DataLayout); - - if (m_Data.m_Parameters.m_BiasEnabled) - { - m_BiasesTensor = std::make_unique<arm_compute::CLTensor>(); - BuildArmComputeTensor(*m_BiasesTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); - } - - m_Data.ValidateInputsOutputs("ClTransposeConvolution2dWorkload", 1, 1); - - arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); - arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); - - arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); - - input.info()->set_data_layout(aclDataLayout); - output.info()->set_data_layout(aclDataLayout); - - arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters); - m_Layer.configure(&input, m_WeightsTensor.get(), m_BiasesTensor.get(), &output, padStrideInfo); - - InitializeArmComputeClTensorData(*m_WeightsTensor, m_Data.m_Weight); - if (m_BiasesTensor) - { - InitializeArmComputeClTensorData(*m_BiasesTensor, m_Data.m_Bias); - } - - m_Layer.prepare(); - - FreeUnusedTensors(); -} - -void ClTransposeConvolution2dWorkload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClTransposeConvolution2dWorkload_Execute"); - RunClFunction(m_Layer, CHECK_LOCATION()); -} - -void ClTransposeConvolution2dWorkload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_WeightsTensor); - FreeTensorIfUnused(m_BiasesTensor); -} - -} // namespace armnn diff --git a/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.hpp b/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.hpp deleted file mode 100644 index b7320bf6ce..0000000000 --- a/src/backends/cl/workloads/ClTransposeConvolution2dWorkload.hpp +++ /dev/null @@ -1,46 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Tensor.hpp> -#include <armnn/Descriptors.hpp> - -#include <backendsCommon/Workload.hpp> - -#include <arm_compute/runtime/CL/functions/CLDeconvolutionLayer.h> -#include <arm_compute/runtime/MemoryManagerOnDemand.h> - -#include <memory> - -namespace armnn -{ - -arm_compute::Status ClTransposeConvolution2dWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const TransposeConvolution2dDescriptor& descriptor, - const TensorInfo& weights, - const Optional<TensorInfo>& biases); - -class ClTransposeConvolution2dWorkload : public BaseWorkload<TransposeConvolution2dQueueDescriptor> -{ -public: - ClTransposeConvolution2dWorkload(const TransposeConvolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info, - std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); - - void Execute() const override; - -private: - mutable arm_compute::CLDeconvolutionLayer m_Layer; - - std::unique_ptr<arm_compute::CLTensor> m_WeightsTensor; - std::unique_ptr<arm_compute::CLTensor> m_BiasesTensor; - - void FreeUnusedTensors(); -}; - -} // namespace armnn - diff --git a/src/backends/cl/workloads/ClWorkloadUtils.hpp b/src/backends/cl/workloads/ClWorkloadUtils.hpp deleted file mode 100644 index b4bcc1c017..0000000000 --- a/src/backends/cl/workloads/ClWorkloadUtils.hpp +++ /dev/null @@ -1,143 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// -#pragma once - -#include <Half.hpp> - -#include <aclCommon/ArmComputeTensorUtils.hpp> -#include <cl/OpenClTimer.hpp> -#include <backendsCommon/CpuTensorHandle.hpp> - -#include <armnn/Utils.hpp> - -#include <arm_compute/runtime/CL/CLTensor.h> -#include <arm_compute/runtime/IFunction.h> - -#include <sstream> - -#define ARMNN_SCOPED_PROFILING_EVENT_CL(name) \ - ARMNN_SCOPED_PROFILING_EVENT_WITH_INSTRUMENTS(armnn::Compute::GpuAcc, \ - name, \ - armnn::OpenClTimer(), \ - armnn::WallClockTimer()) - -namespace armnn -{ - -template <typename T> -void CopyArmComputeClTensorData(arm_compute::CLTensor& dstTensor, const T* srcData) -{ - { - ARMNN_SCOPED_PROFILING_EVENT_CL("MapClTensorForWriting"); - dstTensor.map(true); - } - - { - ARMNN_SCOPED_PROFILING_EVENT_CL("CopyToClTensor"); - armcomputetensorutils::CopyArmComputeITensorData<T>(srcData, dstTensor); - } - - dstTensor.unmap(); -} - -inline auto SetClStridedSliceData(const std::vector<int>& m_begin, - const std::vector<int>& m_end, - const std::vector<int>& m_stride) -{ - arm_compute::Coordinates starts; - arm_compute::Coordinates ends; - arm_compute::Coordinates strides; - - unsigned int num_dims = static_cast<unsigned int>(m_begin.size()); - - for (unsigned int i = 0; i < num_dims; i++) { - unsigned int revertedIndex = num_dims - i - 1; - - starts.set(i, static_cast<int>(m_begin[revertedIndex])); - ends.set(i, static_cast<int>(m_end[revertedIndex])); - strides.set(i, static_cast<int>(m_stride[revertedIndex])); - } - - return std::make_tuple(starts, ends, strides); -} - -inline auto SetClSliceData(const std::vector<unsigned int>& m_begin, - const std::vector<unsigned int>& m_size) -{ - // This function must translate the size vector given to an end vector - // expected by the ACL NESlice workload - arm_compute::Coordinates starts; - arm_compute::Coordinates ends; - - unsigned int num_dims = static_cast<unsigned int>(m_begin.size()); - - // For strided slices, we have the relationship size = (end - begin) / stride - // For slice, we assume stride to be a vector of all ones, yielding the formula - // size = (end - begin) therefore we know end = size + begin - for (unsigned int i = 0; i < num_dims; i++) - { - unsigned int revertedIndex = num_dims - i - 1; - - starts.set(i, static_cast<int>(m_begin[revertedIndex])); - ends.set(i, static_cast<int>(m_begin[revertedIndex] + m_size[revertedIndex])); - } - - return std::make_tuple(starts, ends); -} - -inline void InitializeArmComputeClTensorData(arm_compute::CLTensor& clTensor, - const ConstCpuTensorHandle* handle) -{ - BOOST_ASSERT(handle); - - armcomputetensorutils::InitialiseArmComputeTensorEmpty(clTensor); - switch(handle->GetTensorInfo().GetDataType()) - { - case DataType::Float16: - CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<armnn::Half>()); - break; - case DataType::Float32: - CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<float>()); - break; - case DataType::QAsymmS8: - case DataType::QAsymmU8: - CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<uint8_t>()); - break; - ARMNN_NO_DEPRECATE_WARN_BEGIN - case DataType::QuantizedSymm8PerAxis: - ARMNN_FALLTHROUGH; - case DataType::QSymmS8: - CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<int8_t>()); - break; - ARMNN_NO_DEPRECATE_WARN_END - case DataType::Signed32: - CopyArmComputeClTensorData(clTensor, handle->GetConstTensor<int32_t>()); - break; - default: - BOOST_ASSERT_MSG(false, "Unexpected tensor type."); - } -}; - -inline RuntimeException WrapClError(const cl::Error& clError, const CheckLocation& location) -{ - std::stringstream message; - message << "CL error: " << clError.what() << ". Error code: " << clError.err(); - - return RuntimeException(message.str(), location); -} - -inline void RunClFunction(arm_compute::IFunction& function, const CheckLocation& location) -{ - try - { - function.run(); - } - catch (cl::Error& error) - { - throw WrapClError(error, location); - } -} - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp deleted file mode 100644 index 014dc3f99e..0000000000 --- a/src/backends/cl/workloads/ClWorkloads.hpp +++ /dev/null @@ -1,51 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once -#include "ClAbsWorkload.hpp" -#include "ClActivationWorkload.hpp" -#include "ClAdditionWorkload.hpp" -#include "ClArgMinMaxWorkload.hpp" -#include "ClConstantWorkload.hpp" -#include "ClBatchNormalizationFloatWorkload.hpp" -#include "ClBatchToSpaceNdWorkload.hpp" -#include "ClConvolution2dWorkload.hpp" -#include "ClDepthToSpaceWorkload.hpp" -#include "ClDepthwiseConvolutionWorkload.hpp" -#include "ClDequantizeWorkload.hpp" -#include "ClDivisionFloatWorkload.hpp" -#include "ClFloorFloatWorkload.hpp" -#include "ClFullyConnectedWorkload.hpp" -#include "ClGreaterWorkload.hpp" -#include "ClInstanceNormalizationWorkload.hpp" -#include "ClL2NormalizationFloatWorkload.hpp" -#include "ClLstmFloatWorkload.hpp" -#include "ClConcatWorkload.hpp" -#include "ClMaximumWorkload.hpp" -#include "ClMeanWorkload.hpp" -#include "ClMinimumWorkload.hpp" -#include "ClMultiplicationWorkload.hpp" -#include "ClNormalizationFloatWorkload.hpp" -#include "ClPermuteWorkload.hpp" -#include "ClPadWorkload.hpp" -#include "ClPooling2dWorkload.hpp" -#include "ClPreluWorkload.hpp" -#include "ClQuantizeWorkload.hpp" -#include "ClQuantizedLstmWorkload.hpp" -#include "ClReshapeWorkload.hpp" -#include "ClResizeWorkload.hpp" -#include "ClRsqrtWorkload.hpp" -#include "ClSliceWorkload.hpp" -#include "ClSoftmaxFloatWorkload.hpp" -#include "ClSoftmaxUint8Workload.hpp" -#include "ClSpaceToBatchNdWorkload.hpp" -#include "ClSpaceToDepthWorkload.hpp" -#include "ClSplitterWorkload.hpp" -#include "ClStackWorkload.hpp" -#include "ClStridedSliceWorkload.hpp" -#include "ClSubtractionWorkload.hpp" -#include "ClConvertFp16ToFp32Workload.hpp" -#include "ClConvertFp32ToFp16Workload.hpp" -#include "ClTransposeConvolution2dWorkload.hpp" |