Ryan OShea | 19e7942 | 2022-05-04 00:38:03 +0100 | [diff] [blame] | 1 | // |
| 2 | // Copyright © 2022 Arm Ltd and Contributors. All rights reserved. |
| 3 | // SPDX-License-Identifier: MIT |
| 4 | // |
| 5 | #include "NeonPooling3dWorkload.hpp" |
| 6 | #include "NeonWorkloadUtils.hpp" |
| 7 | #include <neon/NeonLayerSupport.hpp> |
| 8 | #include <neon/NeonTensorHandle.hpp> |
| 9 | #include <aclCommon/ArmComputeUtils.hpp> |
| 10 | #include <aclCommon/ArmComputeTensorUtils.hpp> |
| 11 | |
| 12 | namespace armnn |
| 13 | { |
| 14 | using namespace armcomputetensorutils; |
| 15 | arm_compute::Status NeonPooling3dWorkloadValidate(const TensorInfo& input, |
| 16 | const TensorInfo& output, |
| 17 | const Pooling3dDescriptor& descriptor) |
| 18 | { |
| 19 | const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout); |
| 20 | const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout); |
| 21 | arm_compute::Pooling3dLayerInfo layerInfo = BuildArmComputePooling3dLayerInfo(descriptor); |
| 22 | return arm_compute::NEPooling3dLayer::validate(&aclInputInfo, &aclOutputInfo, layerInfo); |
| 23 | } |
| 24 | |
| 25 | NeonPooling3dWorkload::NeonPooling3dWorkload( const Pooling3dQueueDescriptor& descriptor, |
| 26 | const WorkloadInfo& info) |
| 27 | : NeonBaseWorkload<Pooling3dQueueDescriptor>(descriptor, info) |
| 28 | { |
| 29 | // Report Profiling Details |
| 30 | ARMNN_REPORT_PROFILING_WORKLOAD_DESC("NeonPooling3dWorkload_Construct", |
| 31 | descriptor.m_Parameters, |
| 32 | info, |
| 33 | this->GetGuid()); |
| 34 | |
| 35 | m_Data.ValidateInputsOutputs("NeonPooling3dWorkload", 1, 1); |
| 36 | |
| 37 | arm_compute::ITensor& input = static_cast<IAclTensorHandle*>(m_Data.m_Inputs[0])->GetTensor(); |
| 38 | arm_compute::ITensor& output = static_cast<IAclTensorHandle*>(m_Data.m_Outputs[0])->GetTensor(); |
| 39 | |
| 40 | arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout); |
| 41 | input.info()->set_data_layout(aclDataLayout); |
| 42 | output.info()->set_data_layout(aclDataLayout); |
| 43 | |
| 44 | // flag to use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy |
| 45 | // enable fp_mixed_precision for the the FP16 cases that |
| 46 | // accumulation reaches a limit beyond which there is no more increment of the value |
| 47 | bool fpMixedPrecision = false; |
| 48 | |
| 49 | arm_compute::Pooling3dLayerInfo layerInfo = BuildArmComputePooling3dLayerInfo(m_Data.m_Parameters, |
| 50 | fpMixedPrecision); |
| 51 | { |
| 52 | ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "NeonPooling3dWorkload_configure"); |
| 53 | |
| 54 | auto layer = std::make_unique<arm_compute::NEPooling3dLayer>(); |
| 55 | layer->configure(&input, &output, layerInfo); |
| 56 | m_PoolingLayer.reset(layer.release()); |
| 57 | } |
| 58 | } |
| 59 | void NeonPooling3dWorkload::Execute() const |
| 60 | { |
| 61 | ARMNN_SCOPED_PROFILING_EVENT_NEON_GUID("NeonPooling3dWorkload_Execute", this->GetGuid()); |
| 62 | m_PoolingLayer->run(); |
| 63 | } |
| 64 | } |