diff options
Diffstat (limited to 'src/backends/neon/workloads/NeonStackWorkload.cpp')
-rw-r--r-- | src/backends/neon/workloads/NeonStackWorkload.cpp | 76 |
1 files changed, 76 insertions, 0 deletions
diff --git a/src/backends/neon/workloads/NeonStackWorkload.cpp b/src/backends/neon/workloads/NeonStackWorkload.cpp new file mode 100644 index 0000000000..b21494397d --- /dev/null +++ b/src/backends/neon/workloads/NeonStackWorkload.cpp @@ -0,0 +1,76 @@ +// +// Copyright © 2017 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// +#include "NeonStackWorkload.hpp" +#include "NeonWorkloadUtils.hpp" + +#include <aclCommon/ArmComputeTensorUtils.hpp> +#include <backendsCommon/CpuTensorHandle.hpp> +#include <neon/NeonTensorHandle.hpp> + +#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 NeonStackWorkloadValidate(const std::vector<const TensorInfo*>& inputs, + const TensorInfo& output, + const StackDescriptor& 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); + } + + std::vector<arm_compute::ITensorInfo*> aclInputPtrs; + for (arm_compute::ITensorInfo& input : aclInputs) + { + aclInputPtrs.emplace_back(&input); + } + + const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output); + int aclAxis = CalcAxis(descriptor.m_Axis, descriptor.m_InputShape.GetNumDimensions()); + return arm_compute::NEStackLayer::validate(aclInputPtrs, aclAxis, &aclOutputInfo); +} + +NeonStackWorkload::NeonStackWorkload(const StackQueueDescriptor& descriptor, const WorkloadInfo& info) +: BaseWorkload<StackQueueDescriptor>(descriptor, info) +{ + std::vector<arm_compute::ITensor*> aclInputs; + for (auto input : m_Data.m_Inputs) + { + arm_compute::ITensor& aclInput = boost::polymorphic_pointer_downcast<IAclTensorHandle>(input)->GetTensor(); + aclInputs.emplace_back(&aclInput); + } + arm_compute::ITensor& output = boost::polymorphic_pointer_downcast<IAclTensorHandle>( + m_Data.m_Outputs[0])->GetTensor(); + + m_Layer.reset(new arm_compute::NEStackLayer()); + int aclAxis = CalcAxis(descriptor.m_Parameters.m_Axis, descriptor.m_Parameters.m_InputShape.GetNumDimensions()); + m_Layer->configure(aclInputs, aclAxis, &output); +} + +void NeonStackWorkload::Execute() const +{ + if (m_Layer) + { + ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonStackWorkload_Execute"); + m_Layer->run(); + } +} + +} //namespace armnn |