Teresa Charlin | ec5f7d1 | 2021-10-22 17:15:00 +0100 | [diff] [blame] | 1 | // |
Kevin May | 8eece0a | 2023-06-06 17:19:06 +0100 | [diff] [blame] | 2 | // Copyright © 2021, 2023 Arm Ltd and Contributors. All rights reserved. |
Teresa Charlin | ec5f7d1 | 2021-10-22 17:15:00 +0100 | [diff] [blame] | 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | |
| 6 | #include "NeonConvolution3dWorkload.hpp" |
| 7 | |
| 8 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 9 | #include <aclCommon/ArmComputeUtils.hpp> |
| 10 | #include <armnn/utility/PolymorphicDowncast.hpp> |
Colm Donelan | 0c47974 | 2021-12-10 12:43:54 +0000 | [diff] [blame] | 11 | #include <armnn/backends/TensorHandle.hpp> |
Teresa Charlin | ec5f7d1 | 2021-10-22 17:15:00 +0100 | [diff] [blame] | 12 | #include <neon/workloads/NeonWorkloadUtils.hpp> |
| 13 | |
| 14 | #include <arm_compute/runtime/NEON/functions/NEConv3D.h> |
| 15 | |
| 16 | #include <armnn/Types.hpp> |
| 17 | #include <Half.hpp> |
| 18 | |
| 19 | namespace armnn |
| 20 | { |
| 21 | |
| 22 | using namespace armcomputetensorutils; |
| 23 | |
| 24 | arm_compute::Status NeonConvolution3dWorkloadValidate(const TensorInfo& input, |
| 25 | const TensorInfo& output, |
| 26 | const Convolution3dDescriptor& descriptor, |
| 27 | const TensorInfo& weights, |
| 28 | const Optional<TensorInfo>& biases, |
| 29 | bool isFastMathEnabled, |
| 30 | const ActivationDescriptor* activationDescriptor) |
| 31 | { |
| 32 | const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); |
| 33 | const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout); |
| 34 | arm_compute::TensorInfo aclBiasesInfo; |
| 35 | arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr; |
| 36 | if (descriptor.m_BiasEnabled) |
| 37 | { |
Kevin May | 8eece0a | 2023-06-06 17:19:06 +0100 | [diff] [blame] | 38 | if (!biases.has_value()) |
| 39 | { |
| 40 | return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR, |
| 41 | "ArmNN NeonConvolution3dWorkload has empty bias value."}; |
| 42 | } |
Teresa Charlin | ec5f7d1 | 2021-10-22 17:15:00 +0100 | [diff] [blame] | 43 | |
| 44 | aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout); |
| 45 | optionalAclBiasesInfo = &aclBiasesInfo; |
| 46 | } |
| 47 | const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); |
| 48 | |
| 49 | const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor, |
| 50 | isFastMathEnabled, |
| 51 | activationDescriptor); |
| 52 | |
| 53 | return arm_compute::NEConv3D::validate(&aclInputInfo, |
| 54 | &aclWeightsInfo, |
| 55 | optionalAclBiasesInfo, |
| 56 | &aclOutputInfo, |
| 57 | aclConv3DInfo); |
| 58 | } |
| 59 | |
| 60 | NeonConvolution3dWorkload::NeonConvolution3dWorkload(const Convolution3dQueueDescriptor& descriptor, |
| 61 | const WorkloadInfo& info, |
| 62 | std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager, |
| 63 | const bool isFastMathEnabled) |
Teresa Charlin | 588cbdf | 2022-01-19 15:55:37 +0000 | [diff] [blame] | 64 | : NeonBaseWorkload<Convolution3dQueueDescriptor>(descriptor, info) |
Teresa Charlin | ec5f7d1 | 2021-10-22 17:15:00 +0100 | [diff] [blame] | 65 | { |
| 66 | IgnoreUnused(memoryManager); |
| 67 | |
| 68 | using arm_compute::NEConv3D; |
| 69 | uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2; |
| 70 | m_Data.ValidateInputsOutputs("NeonConvolution3dWorkload", numInputs, 1); |
| 71 | |
| 72 | arm_compute::ITensor& input = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); |
| 73 | arm_compute::ITensor& weights = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Inputs[1])->GetTensor(); |
| 74 | arm_compute::ITensor* biasesPtr = nullptr; |
| 75 | if (m_Data.m_Parameters.m_BiasEnabled) |
| 76 | { |
| 77 | biasesPtr = &PolymorphicDowncast<IAclTensorHandle *>(m_Data.m_Inputs[2])->GetTensor(); |
| 78 | } |
| 79 | arm_compute::ITensor& output = PolymorphicDowncast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); |
| 80 | |
| 81 | arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); |
| 82 | input.info()->set_data_layout(aclDataLayout); |
| 83 | weights.info()->set_data_layout(aclDataLayout); |
| 84 | output.info()->set_data_layout(aclDataLayout); |
| 85 | |
| 86 | const arm_compute::Conv3dInfo aclConv3DInfo = ComputeConv3DInfo(descriptor, isFastMathEnabled); |
| 87 | |
| 88 | auto convolutionLayer = std::make_unique<arm_compute::NEConv3D>(); |
| 89 | convolutionLayer->configure(&input, |
| 90 | &weights, |
| 91 | biasesPtr, |
| 92 | &output, |
| 93 | aclConv3DInfo); |
| 94 | |
| 95 | // Add details for profiling output |
| 96 | WorkloadInfo detailsInfo; |
| 97 | |
| 98 | detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos; |
| 99 | detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos; |
| 100 | |
| 101 | // Report Profiling Details |
| 102 | ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonConvolution3dWorkload_Construct", |
| 103 | descriptor.m_Parameters, |
| 104 | detailsInfo, |
| 105 | this->GetGuid()); |
| 106 | |
| 107 | m_ConvolutionLayer.reset(convolutionLayer.release()); |
| 108 | |
| 109 | ARMNN_ASSERT(m_ConvolutionLayer); |
| 110 | |
| 111 | m_ConvolutionLayer->prepare(); |
| 112 | } |
| 113 | |
| 114 | void NeonConvolution3dWorkload::Execute() const |
| 115 | { |
| 116 | ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonConvolution3dWorkload_Execute", this->GetGuid()); |
| 117 | m_ConvolutionLayer->run(); |
| 118 | } |
| 119 | |
| 120 | } //namespace armnn |