| Derek Lamberti | 6a5e5e8 | 2019-12-05 14:41:20 +0000 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2019 Arm Ltd. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "NeonDetectionPostProcessWorkload.hpp" |
| 7 | |
| 8 | #include "NeonWorkloadUtils.hpp" |
| 9 | |
| 10 | #include <aclCommon/ArmComputeTensorHandle.hpp> |
| 11 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 12 | |
| 13 | #include <boost/cast.hpp> |
| 14 | |
| 15 | namespace armnn |
| 16 | { |
| 17 | |
| 18 | arm_compute::DetectionPostProcessLayerInfo MakeInfo(const DetectionPostProcessDescriptor& desc) |
| 19 | { |
| 20 | return arm_compute::DetectionPostProcessLayerInfo(desc.m_MaxDetections, |
| 21 | desc.m_MaxClassesPerDetection, |
| 22 | desc.m_NmsScoreThreshold, |
| 23 | desc.m_NmsIouThreshold, |
| 24 | desc.m_NumClasses, |
| 25 | { desc.m_ScaleX, |
| 26 | desc.m_ScaleY, |
| 27 | desc.m_ScaleW, |
| 28 | desc.m_ScaleH }, |
| 29 | desc.m_UseRegularNms, |
| 30 | desc.m_DetectionsPerClass); |
| 31 | } |
| 32 | |
| 33 | arm_compute::Status NeonDetectionPostProcessValidate(const TensorInfo& boxEncodings, |
| 34 | const TensorInfo& scores, |
| 35 | const TensorInfo& anchors, |
| 36 | const TensorInfo& detectionBoxes, |
| 37 | const TensorInfo& detectionClasses, |
| 38 | const TensorInfo& detectionScores, |
| 39 | const TensorInfo& numDetections, |
| 40 | const DetectionPostProcessDescriptor &desc) |
| 41 | { |
| 42 | arm_compute::DetectionPostProcessLayerInfo info = MakeInfo(desc); |
| 43 | |
| 44 | const arm_compute::TensorInfo aclBoxEncodings = |
| 45 | armcomputetensorutils::BuildArmComputeTensorInfo(boxEncodings); |
| 46 | |
| 47 | const arm_compute::TensorInfo aclScores = |
| 48 | armcomputetensorutils::BuildArmComputeTensorInfo(scores); |
| 49 | |
| 50 | const arm_compute::TensorInfo aclAnchors = |
| 51 | armcomputetensorutils::BuildArmComputeTensorInfo(anchors); |
| 52 | |
| 53 | arm_compute::TensorInfo aclDetectionBoxes = |
| 54 | armcomputetensorutils::BuildArmComputeTensorInfo(detectionBoxes); |
| 55 | |
| 56 | arm_compute::TensorInfo aclDetectionClasses = |
| 57 | armcomputetensorutils::BuildArmComputeTensorInfo(detectionClasses); |
| 58 | |
| 59 | arm_compute::TensorInfo aclDetectionScores = |
| 60 | armcomputetensorutils::BuildArmComputeTensorInfo(detectionScores); |
| 61 | |
| 62 | arm_compute::TensorInfo aclNumDetections = |
| 63 | armcomputetensorutils::BuildArmComputeTensorInfo(numDetections); |
| 64 | |
| 65 | return arm_compute::CPPDetectionPostProcessLayer::validate( |
| 66 | &aclBoxEncodings, |
| 67 | &aclScores, |
| 68 | &aclAnchors, |
| 69 | &aclDetectionBoxes, |
| 70 | &aclDetectionClasses, |
| 71 | &aclDetectionScores, |
| 72 | &aclNumDetections, |
| 73 | info); |
| 74 | } |
| 75 | |
| 76 | NeonDetectionPostProcessWorkload::NeonDetectionPostProcessWorkload( |
| 77 | const DetectionPostProcessQueueDescriptor& descriptor, |
| 78 | const WorkloadInfo& info) |
| 79 | : BaseWorkload<DetectionPostProcessQueueDescriptor>(descriptor, info) |
| 80 | { |
| 81 | m_Anchors = std::make_unique<arm_compute::Tensor>(); |
| 82 | BuildArmComputeTensor(*m_Anchors, descriptor.m_Anchors->GetTensorInfo()); |
| 83 | |
| 84 | arm_compute::DetectionPostProcessLayerInfo di = MakeInfo(m_Data.m_Parameters); |
| 85 | |
| 86 | auto AclTensorRef = [](ITensorHandle* tensor) -> arm_compute::ITensor& |
| 87 | { |
| 88 | return boost::polymorphic_downcast<IAclTensorHandle*>(tensor)->GetTensor(); |
| 89 | }; |
| 90 | |
| 91 | arm_compute::ITensor& boxEncodings = AclTensorRef(m_Data.m_Inputs[0]); |
| 92 | arm_compute::ITensor& scores = AclTensorRef(m_Data.m_Inputs[1]); |
| 93 | |
| 94 | arm_compute::ITensor& detectionBoxes = AclTensorRef(m_Data.m_Outputs[0]); |
| 95 | arm_compute::ITensor& detectionClasses = AclTensorRef(m_Data.m_Outputs[1]); |
| 96 | arm_compute::ITensor& detectionScores = AclTensorRef(m_Data.m_Outputs[2]); |
| 97 | arm_compute::ITensor& numDetections = AclTensorRef(m_Data.m_Outputs[3]); |
| 98 | |
| 99 | m_Func.configure(&boxEncodings, &scores, m_Anchors.get(), |
| 100 | &detectionBoxes, &detectionClasses, &detectionScores, &numDetections, |
| 101 | di); |
| 102 | |
| 103 | InitializeArmComputeTensorData(*m_Anchors, m_Data.m_Anchors); |
| 104 | } |
| 105 | |
| 106 | void NeonDetectionPostProcessWorkload::Execute() const |
| 107 | { |
| 108 | ARMNN_SCOPED_PROFILING_EVENT_NEON("NeonDetectionPostProcessWorkload_Execute"); |
| 109 | m_Func.run(); |
| 110 | } |
| 111 | |
| 112 | } // namespace armnn |