diff options
Diffstat (limited to 'src/backends/neon/workloads')
4 files changed, 155 insertions, 0 deletions
diff --git a/src/backends/neon/workloads/CMakeLists.txt b/src/backends/neon/workloads/CMakeLists.txt index cf3789e5ac..44db6d263a 100644 --- a/src/backends/neon/workloads/CMakeLists.txt +++ b/src/backends/neon/workloads/CMakeLists.txt @@ -30,6 +30,8 @@ list(APPEND armnnNeonBackendWorkloads_sources NeonDepthwiseConvolutionWorkload.hpp NeonDequantizeWorkload.cpp NeonDequantizeWorkload.hpp + NeonDetectionPostProcessWorkload.cpp + NeonDetectionPostProcessWorkload.hpp NeonFloorFloatWorkload.cpp NeonFloorFloatWorkload.hpp NeonFullyConnectedWorkload.cpp diff --git a/src/backends/neon/workloads/NeonDetectionPostProcessWorkload.cpp b/src/backends/neon/workloads/NeonDetectionPostProcessWorkload.cpp new file mode 100644 index 0000000000..481e950953 --- /dev/null +++ b/src/backends/neon/workloads/NeonDetectionPostProcessWorkload.cpp @@ -0,0 +1,112 @@ +// +// Copyright © 2019 Arm Ltd. All rights reserved. +// SPDX-License-Identifier: MIT +// + +#include "NeonDetectionPostProcessWorkload.hpp" + +#include "NeonWorkloadUtils.hpp" + +#include <aclCommon/ArmComputeTensorHandle.hpp> +#include <aclCommon/ArmComputeTensorUtils.hpp> + +#include <boost/cast.hpp> + +namespace armnn +{ + +arm_compute::DetectionPostProcessLayerInfo MakeInfo(const DetectionPostProcessDescriptor& desc) +{ + return arm_compute::DetectionPostProcessLayerInfo(desc.m_MaxDetections, + desc.m_MaxClassesPerDetection, + desc.m_NmsScoreThreshold, + desc.m_NmsIouThreshold, + desc.m_NumClasses, + { desc.m_ScaleX, + desc.m_ScaleY, + desc.m_ScaleW, + desc.m_ScaleH }, + desc.m_UseRegularNms, + desc.m_DetectionsPerClass); +} + +arm_compute::Status NeonDetectionPostProcessValidate(const TensorInfo& boxEncodings, + const TensorInfo& scores, + const TensorInfo& anchors, + const TensorInfo& detectionBoxes, + const TensorInfo& detectionClasses, + const TensorInfo& detectionScores, + const TensorInfo& numDetections, + const DetectionPostProcessDescriptor &desc) +{ + arm_compute::DetectionPostProcessLayerInfo info = MakeInfo(desc); + + const arm_compute::TensorInfo aclBoxEncodings = + armcomputetensorutils::BuildArmComputeTensorInfo(boxEncodings); + + const arm_compute::TensorInfo aclScores = + armcomputetensorutils::BuildArmComputeTensorInfo(scores); + + const arm_compute::TensorInfo aclAnchors = + armcomputetensorutils::BuildArmComputeTensorInfo(anchors); + + arm_compute::TensorInfo aclDetectionBoxes = + armcomputetensorutils::BuildArmComputeTensorInfo(detectionBoxes); + + arm_compute::TensorInfo aclDetectionClasses = + armcomputetensorutils::BuildArmComputeTensorInfo(detectionClasses); + + arm_compute::TensorInfo aclDetectionScores = + armcomputetensorutils::BuildArmComputeTensorInfo(detectionScores); + + arm_compute::TensorInfo aclNumDetections = + armcomputetensorutils::BuildArmComputeTensorInfo(numDetections); + + return arm_compute::CPPDetectionPostProcessLayer::validate( + &aclBoxEncodings, + &aclScores, + &aclAnchors, + &aclDetectionBoxes, + &aclDetectionClasses, + &aclDetectionScores, + &aclNumDetections, + info); +} + +NeonDetectionPostProcessWorkload::NeonDetectionPostProcessWorkload( + const DetectionPostProcessQueueDescriptor& descriptor, + const WorkloadInfo& info) + : BaseWorkload<DetectionPostProcessQueueDescriptor>(descriptor, info) +{ + m_Anchors = std::make_unique<arm_compute::Tensor>(); + BuildArmComputeTensor(*m_Anchors, descriptor.m_Anchors->GetTensorInfo()); + + arm_compute::DetectionPostProcessLayerInfo di = MakeInfo(m_Data.m_Parameters); + + auto AclTensorRef = [](ITensorHandle* tensor) -> arm_compute::ITensor& + { + return boost::polymorphic_downcast<IAclTensorHandle*>(tensor)->GetTensor(); + }; + + arm_compute::ITensor& boxEncodings = AclTensorRef(m_Data.m_Inputs[0]); + arm_compute::ITensor& scores = AclTensorRef(m_Data.m_Inputs[1]); + + arm_compute::ITensor& detectionBoxes = AclTensorRef(m_Data.m_Outputs[0]); + arm_compute::ITensor& detectionClasses = AclTensorRef(m_Data.m_Outputs[1]); + arm_compute::ITensor& detectionScores = AclTensorRef(m_Data.m_Outputs[2]); + arm_compute::ITensor& numDetections = AclTensorRef(m_Data.m_Outputs[3]); + + m_Func.configure(&boxEncodings, &scores, m_Anchors.get(), + &detectionBoxes, &detectionClasses, &detectionScores, &numDetections, + di); + + InitializeArmComputeTensorData(*m_Anchors, m_Data.m_Anchors); +} + +void NeonDetectionPostProcessWorkload::Execute() const +{ + ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDetectionPostProcessWorkload_Execute"); + m_Func.run(); +} + +} // namespace armnn diff --git a/src/backends/neon/workloads/NeonDetectionPostProcessWorkload.hpp b/src/backends/neon/workloads/NeonDetectionPostProcessWorkload.hpp new file mode 100644 index 0000000000..f7f0ebffb9 --- /dev/null +++ b/src/backends/neon/workloads/NeonDetectionPostProcessWorkload.hpp @@ -0,0 +1,40 @@ +// +// 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/CPP/functions/CPPDetectionPostProcessLayer.h> + +namespace armnn +{ + +arm_compute::Status NeonDetectionPostProcessValidate(const TensorInfo& boxEncodings, + const TensorInfo& scores, + const TensorInfo& anchors, + const TensorInfo& detectionBoxes, + const TensorInfo& detectionClasses, + const TensorInfo& detectionScores, + const TensorInfo& numDetections, + const DetectionPostProcessDescriptor &desc); + +class NeonDetectionPostProcessWorkload : public BaseWorkload<DetectionPostProcessQueueDescriptor> +{ +public: + NeonDetectionPostProcessWorkload( + const DetectionPostProcessQueueDescriptor& descriptor, + const WorkloadInfo& info); + virtual void Execute() const override; + +private: + mutable arm_compute::CPPDetectionPostProcessLayer m_Func; + + std::unique_ptr<arm_compute::Tensor> m_Anchors; + +}; + +} // namespace armnn
\ No newline at end of file diff --git a/src/backends/neon/workloads/NeonWorkloads.hpp b/src/backends/neon/workloads/NeonWorkloads.hpp index 77d819e3ab..dc9bef31b2 100644 --- a/src/backends/neon/workloads/NeonWorkloads.hpp +++ b/src/backends/neon/workloads/NeonWorkloads.hpp @@ -16,6 +16,7 @@ #include "NeonDepthToSpaceWorkload.hpp" #include "NeonDepthwiseConvolutionWorkload.hpp" #include "NeonDequantizeWorkload.hpp" +#include "NeonDetectionPostProcessWorkload.hpp" #include "NeonFloorFloatWorkload.hpp" #include "NeonFullyConnectedWorkload.hpp" #include "NeonGreaterWorkload.hpp" |