diff options
Diffstat (limited to 'src/armnn/backends/NeonWorkloads')
16 files changed, 112 insertions, 56 deletions
diff --git a/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.cpp b/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.cpp index 10c96d82a6..423f02bcb0 100644 --- a/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.cpp @@ -12,9 +12,38 @@ namespace armnn { +using namespace armcomputetensorutils; + +arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const TensorInfo& biases) +{ + const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input); + const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); + const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights); + arm_compute::TensorInfo aclBiasesInfo; + arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; + + if (descriptor.m_BiasEnabled) + { + aclBiasesInfo = BuildArmComputeTensorInfo(biases); + optionalAclBiasesInfo = &aclBiasesInfo; + } + + arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor); + + return arm_compute::NEConvolutionLayer::validate(&aclInputInfo, + &aclWeightsInfo, + optionalAclBiasesInfo, + &aclOutputInfo, + layerInfo); +} + template<armnn::DataType dataType> NeonConvolution2dBaseWorkload<dataType>::NeonConvolution2dBaseWorkload(const Convolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info) + const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) : TypedWorkload<Convolution2dQueueDescriptor, dataType>(descriptor, info) { using arm_compute::NEDirectConvolutionLayer; @@ -50,7 +79,7 @@ NeonConvolution2dBaseWorkload<dataType>::NeonConvolution2dBaseWorkload(const Con if (preferDirectConvolution) { - auto directConvolutionLayer = std::make_unique<arm_compute::NEDirectConvolutionLayer>(); + auto directConvolutionLayer = std::make_unique<arm_compute::NEDirectConvolutionLayer>(memoryManager); directConvolutionLayer->configure(&input, &m_KernelTensor, optionalBiasTensor, @@ -60,7 +89,7 @@ NeonConvolution2dBaseWorkload<dataType>::NeonConvolution2dBaseWorkload(const Con } else { - auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(); + auto convolutionLayer = std::make_unique<arm_compute::NEConvolutionLayer>(memoryManager); convolutionLayer->configure(&input, &m_KernelTensor, optionalBiasTensor, @@ -81,4 +110,3 @@ template class NeonConvolution2dBaseWorkload<DataType::QuantisedAsymm8>; } //namespace armnn - diff --git a/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.hpp b/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.hpp index 98d075a5ea..d28d50d819 100644 --- a/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonConvolution2dBaseWorkload.hpp @@ -12,16 +12,27 @@ #include "backends/ArmComputeTensorUtils.hpp" #include "backends/NeonLayerSupport.hpp" +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + +#include <memory> + namespace armnn { +arm_compute::Status NeonConvolution2dWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const Convolution2dDescriptor& descriptor, + const TensorInfo& weights, + const TensorInfo& biases); + template<armnn::DataType dataType> class NeonConvolution2dBaseWorkload : public TypedWorkload<Convolution2dQueueDescriptor, dataType> { public: using TypedWorkload<Convolution2dQueueDescriptor, dataType>::m_Data; - NeonConvolution2dBaseWorkload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonConvolution2dBaseWorkload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); virtual void ValidateData() const {}; @@ -30,4 +41,5 @@ protected: arm_compute::Tensor m_KernelTensor; arm_compute::Tensor m_BiasTensor; }; -} //namespace armnn
\ No newline at end of file + +} //namespace armnn diff --git a/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.cpp index a8c5c63683..f20f2a4ac5 100644 --- a/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.cpp @@ -13,8 +13,8 @@ namespace armnn using namespace armcomputetensorutils; NeonConvolution2dFloat32Workload::NeonConvolution2dFloat32Workload(const Convolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info) - : NeonConvolution2dBaseWorkload(descriptor, info) + const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) + : NeonConvolution2dBaseWorkload(descriptor, info, memoryManager) { if (m_Data.m_Parameters.m_BiasEnabled) { @@ -22,7 +22,6 @@ NeonConvolution2dFloat32Workload::NeonConvolution2dFloat32Workload(const Convolu } } - void NeonConvolution2dFloat32Workload::Execute() const { ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuAcc, "NeonConvolution2dFloat32Workload_Execute"); @@ -34,8 +33,5 @@ void NeonConvolution2dFloat32Workload::ValidateData() const m_Data.ValidateInputsOutputs("NeonConvolution2dFloat32Workload", 1, 1); } - - } //namespace armnn - diff --git a/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.hpp b/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.hpp index f4d95d623f..56b0848efa 100644 --- a/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonConvolution2dFloat32Workload.hpp @@ -5,21 +5,25 @@ #pragma once -#include <backends/NeonWorkloadUtils.hpp> #include "NeonConvolution2dBaseWorkload.hpp" +#include <backends/NeonWorkloadUtils.hpp> + +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + +#include <memory> namespace armnn { + class NeonConvolution2dFloat32Workload : public NeonConvolution2dBaseWorkload<DataType::Float32> { public: - NeonConvolution2dFloat32Workload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonConvolution2dFloat32Workload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); void Execute() const override; void ValidateData() const override; }; -} //namespace armnn - - +} //namespace armnn diff --git a/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.cpp index ae20522361..fb91f7b7b2 100644 --- a/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.cpp @@ -5,12 +5,12 @@ #include "NeonConvolution2dUint8Workload.hpp" - namespace armnn { + NeonConvolution2dUint8Workload::NeonConvolution2dUint8Workload(const Convolution2dQueueDescriptor& descriptor, - const WorkloadInfo& info) - : NeonConvolution2dBaseWorkload(descriptor, info) + const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) + : NeonConvolution2dBaseWorkload(descriptor, info, memoryManager) { if (m_Data.m_Parameters.m_BiasEnabled) { @@ -21,7 +21,7 @@ NeonConvolution2dUint8Workload::NeonConvolution2dUint8Workload(const Convolution void NeonConvolution2dUint8Workload::Execute() const { - ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuAcc, NeonConvolution2dUint8Workload_Execute); + ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuAcc, "NeonConvolution2dUint8Workload_Execute"); m_ConvolutionLayer->run(); } @@ -30,4 +30,4 @@ void NeonConvolution2dUint8Workload::ValidateData() const m_Data.ValidateInputsOutputs("NeonConvolution2dUint8Workload", 1, 1); } -} //namespace armnn
\ No newline at end of file +} //namespace armnn diff --git a/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.hpp b/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.hpp index 319d574b1e..5b977210c4 100644 --- a/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonConvolution2dUint8Workload.hpp @@ -7,13 +7,18 @@ #include "NeonConvolution2dBaseWorkload.hpp" +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + +#include <memory> + namespace armnn { class NeonConvolution2dUint8Workload : public NeonConvolution2dBaseWorkload<DataType::QuantisedAsymm8> { public: - NeonConvolution2dUint8Workload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonConvolution2dUint8Workload(const Convolution2dQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); virtual void ValidateData() const override; virtual void Execute() const override; @@ -22,6 +27,3 @@ private: } //namespace armnnn - - - diff --git a/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.cpp index 54c4e4333c..e1c4448642 100644 --- a/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.cpp @@ -7,14 +7,14 @@ #include "backends/CpuTensorHandle.hpp" #include "backends/ArmComputeTensorUtils.hpp" - namespace armnn { using namespace armcomputetensorutils; NeonFullyConnectedFloat32Workload::NeonFullyConnectedFloat32Workload(const FullyConnectedQueueDescriptor& descriptor, - const WorkloadInfo& info) + const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) : Float32Workload<FullyConnectedQueueDescriptor>(descriptor, info) + , m_FullyConnectedLayer(memoryManager) { m_Data.ValidateInputsOutputs("NeonFullyConnectedFloat32Workload", 1, 1); @@ -51,4 +51,3 @@ void NeonFullyConnectedFloat32Workload::Execute() const } //namespace armnn - diff --git a/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.hpp b/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.hpp index f9230f1d93..9c722dc573 100644 --- a/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonFullyConnectedFloat32Workload.hpp @@ -7,13 +7,18 @@ #include <backends/NeonWorkloadUtils.hpp> +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + +#include <memory> + namespace armnn { class NeonFullyConnectedFloat32Workload : public Float32Workload<FullyConnectedQueueDescriptor> { public: - NeonFullyConnectedFloat32Workload(const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonFullyConnectedFloat32Workload(const FullyConnectedQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); virtual void Execute() const override; private: @@ -24,7 +29,3 @@ private: } //namespace armnn - - - - diff --git a/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.cpp index 085f58a219..9f79fa09de 100644 --- a/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.cpp @@ -6,13 +6,13 @@ #include "NeonL2NormalizationFloat32Workload.hpp" #include "backends/ArmComputeUtils.hpp" - namespace armnn { NeonL2NormalizationFloat32Workload::NeonL2NormalizationFloat32Workload(const L2NormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) + const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) : Float32Workload<L2NormalizationQueueDescriptor>(descriptor, info) + , m_Layer(memoryManager) { m_Data.ValidateInputsOutputs("NeonL2NormalizationFloat32Workload", 1, 1); diff --git a/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.hpp b/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.hpp index 6cab28366a..2b4a1fef37 100644 --- a/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonL2NormalizationFloat32Workload.hpp @@ -7,20 +7,24 @@ #include <backends/NeonWorkloadUtils.hpp> +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + +#include <memory> + namespace armnn { + class NeonL2NormalizationFloat32Workload : public Float32Workload<L2NormalizationQueueDescriptor> { public: - NeonL2NormalizationFloat32Workload(const L2NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonL2NormalizationFloat32Workload(const L2NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); virtual void Execute() const override; private: // Purposely not a NEL2Normalize function. See constructor. mutable arm_compute::NENormalizationLayer m_Layer; }; -} //namespace armnn - - +} //namespace armnn diff --git a/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.cpp index 739390d5a1..0fd0dcc420 100644 --- a/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.cpp @@ -11,8 +11,9 @@ namespace armnn { NeonNormalizationFloat32Workload::NeonNormalizationFloat32Workload(const NormalizationQueueDescriptor& descriptor, - const WorkloadInfo& info) + const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) : Float32Workload<NormalizationQueueDescriptor>(descriptor, info) + , m_NormalizationLayer(memoryManager) { m_Data.ValidateInputsOutputs("NeonNormalizationFloat32Workload", 1, 1); std::string reasonIfUnsupported; diff --git a/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.hpp b/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.hpp index 12a0fa80b2..24b6da8528 100644 --- a/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonNormalizationFloat32Workload.hpp @@ -7,13 +7,16 @@ #include <backends/NeonWorkloadUtils.hpp> +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + namespace armnn { class NeonNormalizationFloat32Workload : public Float32Workload<NormalizationQueueDescriptor> { public: - NeonNormalizationFloat32Workload(const NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonNormalizationFloat32Workload(const NormalizationQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); virtual void Execute() const override; private: diff --git a/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.cpp index 229562ece2..5e2925ca02 100644 --- a/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.cpp @@ -7,9 +7,11 @@ namespace armnn { + NeonSoftmaxFloat32Workload::NeonSoftmaxFloat32Workload(const SoftmaxQueueDescriptor& descriptor, - const WorkloadInfo& info) + const WorkloadInfo& info, std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) : Float32Workload<SoftmaxQueueDescriptor>(descriptor, info) + , m_SoftmaxLayer(memoryManager) { m_Data.ValidateInputsOutputs("NeonSoftmaxFloat32Workload", 1, 1); @@ -25,7 +27,6 @@ void NeonSoftmaxFloat32Workload::Execute() const ARMNN_SCOPED_PROFILING_EVENT(Compute::CpuAcc, "NeonSoftmaxFloat32Workload_Execute"); m_SoftmaxLayer.run(); } -} //namespace armnn - +} //namespace armnn diff --git a/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.hpp b/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.hpp index c466a0f9c6..91d25b47f8 100644 --- a/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonSoftmaxFloat32Workload.hpp @@ -7,13 +7,18 @@ #include <backends/NeonWorkloadUtils.hpp> +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + +#include <memory> + namespace armnn { class NeonSoftmaxFloat32Workload : public Float32Workload<SoftmaxQueueDescriptor> { public: - NeonSoftmaxFloat32Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonSoftmaxFloat32Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); virtual void Execute() const override; private: @@ -22,6 +27,3 @@ private: } //namespace armnn - - - diff --git a/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.cpp b/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.cpp index a66b0343ff..eb4a23c13c 100644 --- a/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.cpp +++ b/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.cpp @@ -5,12 +5,14 @@ #include "NeonSoftmaxUint8Workload.hpp" - - namespace armnn { -NeonSoftmaxUint8Workload::NeonSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info) + +NeonSoftmaxUint8Workload::NeonSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, + const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager) : Uint8Workload<SoftmaxQueueDescriptor>(descriptor, info) + , m_SoftmaxLayer(memoryManager) { m_Data.ValidateInputsOutputs("NeonSoftmaxUint8Workload", 1, 1); @@ -34,5 +36,6 @@ void NeonSoftmaxUint8Workload::Execute() const m_SoftmaxLayer.run(); } + } //namespace armnn diff --git a/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.hpp b/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.hpp index bccd82a850..19549ef3ef 100644 --- a/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.hpp +++ b/src/armnn/backends/NeonWorkloads/NeonSoftmaxUint8Workload.hpp @@ -7,13 +7,16 @@ #include <backends/NeonWorkloadUtils.hpp> +#include "arm_compute/runtime/MemoryManagerOnDemand.h" + namespace armnn { class NeonSoftmaxUint8Workload : public Uint8Workload<SoftmaxQueueDescriptor> { public: - NeonSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info); + NeonSoftmaxUint8Workload(const SoftmaxQueueDescriptor& descriptor, const WorkloadInfo& info, + std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager); virtual void Execute() const override; private: @@ -22,6 +25,3 @@ private: } //namespace armnn - - - |