diff options
Diffstat (limited to 'src/backends/cl')
-rw-r--r-- | src/backends/cl/ClLayerSupport.cpp | 2 | ||||
-rw-r--r-- | src/backends/cl/ClWorkloadFactory.cpp | 3 | ||||
-rw-r--r-- | src/backends/cl/backend.mk | 4 | ||||
-rw-r--r-- | src/backends/cl/workloads/CMakeLists.txt | 8 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClConvolution2dBaseWorkload.cpp | 48 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClConvolution2dBaseWorkload.hpp | 24 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClConvolution2dFloatWorkload.hpp | 35 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClConvolution2dUint8Workload.cpp | 85 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClConvolution2dUint8Workload.hpp | 35 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClConvolution2dWorkload.cpp (renamed from src/backends/cl/workloads/ClConvolution2dFloatWorkload.cpp) | 56 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClConvolution2dWorkload.hpp | 46 | ||||
-rw-r--r-- | src/backends/cl/workloads/ClWorkloads.hpp | 3 |
12 files changed, 97 insertions, 252 deletions
diff --git a/src/backends/cl/ClLayerSupport.cpp b/src/backends/cl/ClLayerSupport.cpp index a17997b184..e23c70ec30 100644 --- a/src/backends/cl/ClLayerSupport.cpp +++ b/src/backends/cl/ClLayerSupport.cpp @@ -19,7 +19,7 @@ #include "workloads/ClBatchNormalizationFloatWorkload.hpp" #include "workloads/ClConvertFp16ToFp32Workload.hpp" #include "workloads/ClConvertFp32ToFp16Workload.hpp" -#include "workloads/ClConvolution2dBaseWorkload.hpp" +#include "workloads/ClConvolution2dWorkload.hpp" #include "workloads/ClDepthwiseConvolutionBaseWorkload.hpp" #include "workloads/ClDivisionFloatWorkload.hpp" #include "workloads/ClL2NormalizationFloatWorkload.hpp" diff --git a/src/backends/cl/ClWorkloadFactory.cpp b/src/backends/cl/ClWorkloadFactory.cpp index bf17382031..2e074450dd 100644 --- a/src/backends/cl/ClWorkloadFactory.cpp +++ b/src/backends/cl/ClWorkloadFactory.cpp @@ -144,8 +144,7 @@ std::unique_ptr<armnn::IWorkload> ClWorkloadFactory::CreatePooling2d(const Pooli std::unique_ptr<armnn::IWorkload> ClWorkloadFactory::CreateConvolution2d(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info) const { - return MakeWorkload<ClConvolution2dFloatWorkload, ClConvolution2dUint8Workload>(descriptor, info, - m_MemoryManager.GetIntraLayerManager()); + return std::make_unique<ClConvolution2dWorkload>(descriptor, info, m_MemoryManager.GetIntraLayerManager()); } std::unique_ptr<IWorkload> ClWorkloadFactory::CreateDepthwiseConvolution2d( diff --git a/src/backends/cl/backend.mk b/src/backends/cl/backend.mk index 81540cc251..c86940cda1 100644 --- a/src/backends/cl/backend.mk +++ b/src/backends/cl/backend.mk @@ -21,9 +21,7 @@ BACKEND_SOURCES := \ workloads/ClConstantUint8Workload.cpp \ workloads/ClConvertFp16ToFp32Workload.cpp \ workloads/ClConvertFp32ToFp16Workload.cpp \ - workloads/ClConvolution2dBaseWorkload.cpp \ - workloads/ClConvolution2dFloatWorkload.cpp \ - workloads/ClConvolution2dUint8Workload.cpp \ + workloads/ClConvolution2dWorkload.cpp \ workloads/ClDepthwiseConvolutionBaseWorkload.cpp \ workloads/ClDepthwiseConvolutionFloatWorkload.cpp \ workloads/ClDepthwiseConvolutionUint8Workload.cpp \ diff --git a/src/backends/cl/workloads/CMakeLists.txt b/src/backends/cl/workloads/CMakeLists.txt index 9c48dc9e50..5a1653e8e3 100644 --- a/src/backends/cl/workloads/CMakeLists.txt +++ b/src/backends/cl/workloads/CMakeLists.txt @@ -23,12 +23,8 @@ list(APPEND armnnClBackendWorkloads_sources ClConvertFp16ToFp32Workload.hpp ClConvertFp32ToFp16Workload.cpp ClConvertFp32ToFp16Workload.hpp - ClConvolution2dBaseWorkload.cpp - ClConvolution2dBaseWorkload.hpp - ClConvolution2dFloatWorkload.cpp - ClConvolution2dFloatWorkload.hpp - ClConvolution2dUint8Workload.cpp - ClConvolution2dUint8Workload.hpp + ClConvolution2dWorkload.cpp + ClConvolution2dWorkload.hpp ClDepthwiseConvolutionBaseWorkload.cpp ClDepthwiseConvolutionBaseWorkload.hpp ClDepthwiseConvolutionFloatWorkload.cpp diff --git a/src/backends/cl/workloads/ClConvolution2dBaseWorkload.cpp b/src/backends/cl/workloads/ClConvolution2dBaseWorkload.cpp deleted file mode 100644 index 58699a8287..0000000000 --- a/src/backends/cl/workloads/ClConvolution2dBaseWorkload.cpp +++ /dev/null @@ -1,48 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClConvolution2dBaseWorkload.hpp" -#include <backends/cl/ClLayerSupport.hpp> -#include <backends/cl/ClTensorHandle.hpp> -#include <backends/aclCommon/ArmComputeUtils.hpp> -#include <backends/aclCommon/ArmComputeTensorUtils.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 boost::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); - - arm_compute::TensorInfo aclBiasesInfo; - arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; - - if (descriptor.m_BiasEnabled) - { - BOOST_ASSERT(biases.is_initialized()); - - aclBiasesInfo = BuildArmComputeTensorInfo(biases.get(), descriptor.m_DataLayout); - optionalAclBiasesInfo = &aclBiasesInfo; - } - - arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor); - - return arm_compute::CLConvolutionLayer::validate(&aclInputInfo, - &aclWeightsInfo, - optionalAclBiasesInfo, - &aclOutputInfo, - layerInfo); -} - -} diff --git a/src/backends/cl/workloads/ClConvolution2dBaseWorkload.hpp b/src/backends/cl/workloads/ClConvolution2dBaseWorkload.hpp deleted file mode 100644 index a983dba79a..0000000000 --- a/src/backends/cl/workloads/ClConvolution2dBaseWorkload.hpp +++ /dev/null @@ -1,24 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <armnn/Tensor.hpp> -#include <armnn/Descriptors.hpp> - -#include <boost/optional.hpp> - -#include <arm_compute/core/Error.h> - -namespace armnn -{ - -arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input, - const TensorInfo& output, - const Convolution2dDescriptor& descriptor, - const TensorInfo& weights, - const boost::optional<TensorInfo>& biases); - -} //namespace armnn diff --git a/src/backends/cl/workloads/ClConvolution2dFloatWorkload.hpp b/src/backends/cl/workloads/ClConvolution2dFloatWorkload.hpp deleted file mode 100644 index 1f9710e1ea..0000000000 --- a/src/backends/cl/workloads/ClConvolution2dFloatWorkload.hpp +++ /dev/null @@ -1,35 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backends/Workload.hpp> - -#include <arm_compute/runtime/CL/CLFunctions.h> -#include <arm_compute/runtime/MemoryManagerOnDemand.h> - -#include <memory> - -namespace armnn -{ - -class ClConvolution2dFloatWorkload : public FloatWorkload<Convolution2dQueueDescriptor> -{ -public: - ClConvolution2dFloatWorkload(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/ClConvolution2dUint8Workload.cpp b/src/backends/cl/workloads/ClConvolution2dUint8Workload.cpp deleted file mode 100644 index 4f8da34e31..0000000000 --- a/src/backends/cl/workloads/ClConvolution2dUint8Workload.cpp +++ /dev/null @@ -1,85 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#include "ClConvolution2dUint8Workload.hpp" -#include <backends/cl/ClTensorHandle.hpp> -#include <backends/CpuTensorHandle.hpp> -#include <backends/aclCommon/ArmComputeTensorUtils.hpp> -#include <backends/cl/ClLayerSupport.hpp> - -#include "ClWorkloadUtils.hpp" - -namespace armnn -{ -using namespace armcomputetensorutils; - -ClConvolution2dUint8Workload::ClConvolution2dUint8Workload(const Convolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) - : Uint8Workload<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); - - arm_compute::PadStrideInfo padStrideInfo(m_Data.m_Parameters.m_StrideX, - m_Data.m_Parameters.m_StrideY, - m_Data.m_Parameters.m_PadLeft, - m_Data.m_Parameters.m_PadRight, - m_Data.m_Parameters.m_PadTop, - m_Data.m_Parameters.m_PadBottom, - arm_compute::DimensionRoundingType::FLOOR); - - 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("ClConvolution2dUint8Workload", 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_ConvolutionLayer.configure(&input, - m_KernelTensor.get(), - m_BiasTensor.get(), - &output, - padStrideInfo); - - 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 ClConvolution2dUint8Workload::Execute() const -{ - ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvolution2dUint8Workload_Execute"); - - m_ConvolutionLayer.run(); -} - -void ClConvolution2dUint8Workload::FreeUnusedTensors() -{ - FreeTensorIfUnused(m_KernelTensor); - FreeTensorIfUnused(m_BiasTensor); -} - -} //namespace armnn - diff --git a/src/backends/cl/workloads/ClConvolution2dUint8Workload.hpp b/src/backends/cl/workloads/ClConvolution2dUint8Workload.hpp deleted file mode 100644 index 1720ec935c..0000000000 --- a/src/backends/cl/workloads/ClConvolution2dUint8Workload.hpp +++ /dev/null @@ -1,35 +0,0 @@ -// -// Copyright © 2017 Arm Ltd. All rights reserved. -// SPDX-License-Identifier: MIT -// - -#pragma once - -#include <backends/Workload.hpp> - -#include <arm_compute/runtime/CL/CLFunctions.h> -#include <arm_compute/runtime/MemoryManagerOnDemand.h> - -#include <memory> - -namespace armnn -{ - -class ClConvolution2dUint8Workload : public Uint8Workload<Convolution2dQueueDescriptor> -{ -public: - ClConvolution2dUint8Workload(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/ClConvolution2dFloatWorkload.cpp b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp index 0d70ddbd6c..521711becc 100644 --- a/src/backends/cl/workloads/ClConvolution2dFloatWorkload.cpp +++ b/src/backends/cl/workloads/ClConvolution2dWorkload.cpp @@ -3,24 +3,58 @@ // SPDX-License-Identifier: MIT // -#include "ClConvolution2dFloatWorkload.hpp" +#include "ClConvolution2dWorkload.hpp" + +#include "ClWorkloadUtils.hpp" + +#include <backends/cl/ClLayerSupport.hpp> #include <backends/cl/ClTensorHandle.hpp> -#include <backends/CpuTensorHandle.hpp> -#include <backends/aclCommon/ArmComputeTensorUtils.hpp> #include <backends/cl/ClLayerSupport.hpp> +#include <backends/aclCommon/ArmComputeUtils.hpp> +#include <backends/aclCommon/ArmComputeTensorUtils.hpp> +#include <backends/CpuTensorHandle.hpp> -#include "ClWorkloadUtils.hpp" +#include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h> namespace armnn { using namespace armcomputetensorutils; -ClConvolution2dFloatWorkload::ClConvolution2dFloatWorkload(const Convolution2dQueueDescriptor& descriptor, +arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const boost::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); + + arm_compute::TensorInfo aclBiasesInfo; + arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; + + if (descriptor.m_BiasEnabled) + { + BOOST_ASSERT(biases.is_initialized()); + + aclBiasesInfo = BuildArmComputeTensorInfo(biases.get(), descriptor.m_DataLayout); + optionalAclBiasesInfo = &aclBiasesInfo; + } + + arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor); + + return arm_compute::CLConvolutionLayer::validate(&aclInputInfo, + &aclWeightsInfo, + optionalAclBiasesInfo, + &aclOutputInfo, + layerInfo); +} + +ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) - : FloatWorkload<Convolution2dQueueDescriptor>(descriptor, info) + : BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info) , m_ConvolutionLayer(memoryManager) { - // todo: check tensor shapes match. const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo(); @@ -41,7 +75,7 @@ ClConvolution2dFloatWorkload::ClConvolution2dFloatWorkload(const Convolution2dQu BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout); } - m_Data.ValidateInputsOutputs("ClConvolution2dFloat32Workload", 1, 1); + 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(); @@ -69,14 +103,14 @@ ClConvolution2dFloatWorkload::ClConvolution2dFloatWorkload(const Convolution2dQu FreeUnusedTensors(); } -void ClConvolution2dFloatWorkload::Execute() const +void ClConvolution2dWorkload::Execute() const { - ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvolution2dFloat32Workload_Execute"); + ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvolution2dWorkload_Execute"); m_ConvolutionLayer.run(); } -void ClConvolution2dFloatWorkload::FreeUnusedTensors() +void ClConvolution2dWorkload::FreeUnusedTensors() { FreeTensorIfUnused(m_KernelTensor); FreeTensorIfUnused(m_BiasTensor); diff --git a/src/backends/cl/workloads/ClConvolution2dWorkload.hpp b/src/backends/cl/workloads/ClConvolution2dWorkload.hpp new file mode 100644 index 0000000000..14a39f3b25 --- /dev/null +++ b/src/backends/cl/workloads/ClConvolution2dWorkload.hpp @@ -0,0 +1,46 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include <armnn/Tensor.hpp> +#include <armnn/Descriptors.hpp> + +#include <backends/Workload.hpp> + +#include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h> +#include <arm_compute/runtime/MemoryManagerOnDemand.h> + +#include <boost/optional.hpp> + +#include <memory> + +namespace armnn +{ + +arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const boost::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/ClWorkloads.hpp b/src/backends/cl/workloads/ClWorkloads.hpp index 439ad2d989..0ef8fd34bf 100644 --- a/src/backends/cl/workloads/ClWorkloads.hpp +++ b/src/backends/cl/workloads/ClWorkloads.hpp @@ -12,8 +12,7 @@ #include "ClBatchNormalizationFloatWorkload.hpp" #include "ClConstantFloatWorkload.hpp" #include "ClConstantUint8Workload.hpp" -#include "ClConvolution2dFloatWorkload.hpp" -#include "ClConvolution2dUint8Workload.hpp" +#include "ClConvolution2dWorkload.hpp" #include "ClDepthwiseConvolutionFloatWorkload.hpp" #include "ClDepthwiseConvolutionUint8Workload.hpp" #include "ClDivisionFloatWorkload.hpp" |