1 //
2 // Copyright © 2017 Arm Ltd and Contributors. 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,const arm_compute::CLCompileContext & clCompileContext)30 ClPooling2dWorkload::ClPooling2dWorkload(
31 const Pooling2dQueueDescriptor& descriptor,
32 const WorkloadInfo& info,
33 const arm_compute::CLCompileContext& clCompileContext)
34 : ClBaseWorkload<Pooling2dQueueDescriptor>(descriptor, info)
35 {
36 // Report Profiling Details
37 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClPooling2dWorkload_Construct",
38 descriptor.m_Parameters,
39 info,
40 this->GetGuid());
41
42 m_Data.ValidateInputsOutputs("ClPooling2dWorkload", 1, 1);
43
44 arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
45 arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
46
47 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
48 input.info()->set_data_layout(aclDataLayout);
49 output.info()->set_data_layout(aclDataLayout);
50
51 // flag to use wider accumulators (32 bit instead of 16 for FP16) to improve accuracy
52 // enable fp_mixed_precision for the the FP16 cases that
53 // accumulation reaches a limit beyond which there is no more increment of the value
54 bool fpMixedPrecision = false;
55
56 arm_compute::PoolingLayerInfo layerInfo = BuildArmComputePoolingLayerInfo(m_Data.m_Parameters, fpMixedPrecision);
57
58 {
59 ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClPooling2dWorkload_configure");
60 // Run the layer.
61 m_PoolingLayer.configure(clCompileContext, &input, &output, layerInfo);
62 }
63 }
64
Execute() const65 void ClPooling2dWorkload::Execute() const
66 {
67 ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClPooling2dWorkload_Execute", this->GetGuid());
68 RunClFunction(m_PoolingLayer, CHECK_LOCATION());
69 }
70
71 }
72