diff options
Diffstat (limited to 'src/backends/neon/workloads')
-rw-r--r-- | src/backends/neon/workloads/CMakeLists.txt | 2 | ||||
-rw-r--r-- | src/backends/neon/workloads/NeonPooling3dWorkload.cpp | 64 | ||||
-rw-r--r-- | src/backends/neon/workloads/NeonPooling3dWorkload.hpp | 33 | ||||
-rw-r--r-- | src/backends/neon/workloads/NeonWorkloads.hpp | 1 |
4 files changed, 100 insertions, 0 deletions
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt index 16fae69029..2209bf47e2 100644 --- a/src/backends/neon/workloads/CMakeLists.txt +++ b/src/backends/neon/workloads/CMakeLists.txt @@ -94,6 +94,8 @@ list(APPEND armnnNeonBackendWorkloads_sources NeonPermuteWorkload.hpp NeonPooling2dWorkload.cpp NeonPooling2dWorkload.hpp + NeonPooling3dWorkload.hpp + NeonPooling3dWorkload.cpp NeonPreluWorkload.cpp NeonPreluWorkload.hpp NeonQLstmWorkload.cpp diff --git a/src/backends/neon/workloads/NeonPooling3dWorkload.cpp b/src/backends/neon/workloads/NeonPooling3dWorkload.cpp new file mode 100644 index 0000000000..4bc4f986c3 --- /dev/null +++ b/src/backends/neon/workloads/NeonPooling3dWorkload.cpp @@ -0,0 +1,64 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// +#include "NeonPooling3dWorkload.hpp" +#include "NeonWorkloadUtils.hpp" +#include <neon/NeonLayerSupport.hpp> +#include <neon/NeonTensorHandle.hpp> +#include <aclCommon/ArmComputeUtils.hpp> +#include <aclCommon/ArmComputeTensorUtils.hpp> + +namespace armnn +{ + using namespace armcomputetensorutils; + arm_compute::Status NeonPooling3dWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const Pooling3dDescriptor& descriptor) + { + const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); + const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); + arm_compute::Pooling3dLayerInfo layerInfo = BuildArmComputePooling3dLayerInfo(descriptor); + return arm_compute::NEPooling3dLayer::validate(&aclInputInfo, &aclOutputInfo, layerInfo); + } + + NeonPooling3dWorkload::NeonPooling3dWorkload( const Pooling3dQueueDescriptor& descriptor, + const WorkloadInfo& info) + : NeonBaseWorkload<Pooling3dQueueDescriptor>(descriptor, info) + { + // Report Profiling Details + ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonPooling3dWorkload_Construct", + descriptor.m_Parameters, + info, + this->GetGuid()); + + m_Data.ValidateInputsOutputs("NeonPooling3dWorkload", 1, 1); + + arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); + arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(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::Pooling3dLayerInfo layerInfo = BuildArmComputePooling3dLayerInfo(m_Data.m_Parameters, + fpMixedPrecision); + { + ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "NeonPooling3dWorkload_configure"); + + auto layer = std::make_unique<arm_compute::NEPooling3dLayer>(); + layer->configure(&input, &output, layerInfo); + m_PoolingLayer.reset(layer.release()); + } + } + void NeonPooling3dWorkload::Execute() const + { + ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonPooling3dWorkload_Execute", this->GetGuid()); + m_PoolingLayer->run(); + } +} diff --git a/src/backends/neon/workloads/NeonPooling3dWorkload.hpp b/src/backends/neon/workloads/NeonPooling3dWorkload.hpp new file mode 100644 index 0000000000..03453f1eee --- /dev/null +++ b/src/backends/neon/workloads/NeonPooling3dWorkload.hpp @@ -0,0 +1,33 @@ +// +// Copyright © 2022 Arm Ltd and Contributors. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#pragma once + +#include "NeonBaseWorkload.hpp" + +#include <arm_compute/runtime/NEON/functions/NEPooling3dLayer.h> + +namespace armnn +{ + + arm_compute::Status NeonPooling3dWorkloadValidate(const TensorInfo& input, + const TensorInfo& output, + const Pooling3dDescriptor& descriptor); + + class NeonPooling3dWorkload : public NeonBaseWorkload<Pooling3dQueueDescriptor> + { + public: + using BaseWorkload<Pooling3dQueueDescriptor>::m_Data; + + NeonPooling3dWorkload(const Pooling3dQueueDescriptor& descriptor, + const WorkloadInfo& info); + + void Execute() const override; + + private: + std::unique_ptr<arm_compute::IFunction> m_PoolingLayer; + }; + +} //namespace armnn diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp index d486a1d988..8f83674950 100644 --- a/src/backends/neon/workloads/NeonWorkloads.hpp +++ b/src/backends/neon/workloads/NeonWorkloads.hpp @@ -49,6 +49,7 @@ #include "NeonPadWorkload.hpp" #include "NeonPermuteWorkload.hpp" #include "NeonPooling2dWorkload.hpp" +#include "NeonPooling3dWorkload.hpp" #include "NeonPreluWorkload.hpp" #include "NeonQLstmWorkload.hpp" #include "NeonQuantizedLstmWorkload.hpp" |