1 //
2 // Copyright © 2017 Arm Ltd. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "ClPooling2dWorkload.hpp"
7 #include <cl/ClLayerSupport.hpp>
8 #include <cl/ClTensorHandle.hpp>
9 #include <aclCommon/ArmComputeUtils.hpp>
10 #include <aclCommon/ArmComputeTensorUtils.hpp>
11
12 #include "ClWorkloadUtils.hpp"
13
14 namespace armnn
15 {
16 using namespace armcomputetensorutils;
17
ClPooling2dWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const Pooling2dDescriptor & descriptor)18 arm_compute::Status ClPooling2dWorkloadValidate(const TensorInfo& input,
19 const TensorInfo& output,
20 const Pooling2dDescriptor& descriptor)
21 {
22 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
23 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
24
25 arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(descriptor);
26
27 return arm_compute::CLPoolingLayer::validate(&aclInputInfo, &aclOutputInfo, layerInfo);
28 }
29
ClPooling2dWorkload(const Pooling2dQueueDescriptor & descriptor,const WorkloadInfo & info)30 ClPooling2dWorkload::ClPooling2dWorkload(
31 const Pooling2dQueueDescriptor& descriptor, const WorkloadInfo& info)
32 : BaseWorkload<Pooling2dQueueDescriptor>(descriptor, info)
33 {
34 m_Data.ValidateInputsOutputs("ClPooling2dWorkload", 1, 1);
35
36 arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
37 arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
38
39 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
40 input.info()->set_data_layout(aclDataLayout);
41 output.info()->set_data_layout(aclDataLayout);
42
43 // flag to use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
44 // enable fp_mixed_precision for the the FP16 cases that
45 // accumulation reaches a limit beyond which there is no more increment of the value
46 bool fpMixedPrecision = false;
47
48 arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(m_Data.m_Parameters, fpMixedPrecision);
49
50 // Run the layer.
51 m_PoolingLayer.configure(&input, &output, layerInfo);
52 }
53
Execute() const54 void ClPooling2dWorkload::Execute() const
55 {
56 ARMNN_SCOPED_PROFILING_EVENT_CL("ClPooling2dWorkload_Execute");
57 RunClFunction(m_PoolingLayer, CHECK_LOCATION());
58 }
59
60 }
61