1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #include "ClConvolution2dWorkload.hpp"
7
8 #include "ClWorkloadUtils.hpp"
9
10 #include <cl/ClLayerSupport.hpp>
11 #include <cl/ClTensorHandle.hpp>
12 #include <cl/ClLayerSupport.hpp>
13 #include <aclCommon/ArmComputeUtils.hpp>
14 #include <aclCommon/ArmComputeTensorUtils.hpp>
15 #include <armnn/backends/TensorHandle.hpp>
16
17 #include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h>
18
19 namespace armnn
20 {
21 using namespace armcomputetensorutils;
22
ClConvolution2dWorkloadValidate(const TensorInfo & input,const TensorInfo & output,const Convolution2dDescriptor & descriptor,const TensorInfo & weights,const Optional<TensorInfo> & biases,bool isFastMathEnabled,const ActivationDescriptor * activationDescriptor)23 arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
24 const TensorInfo& output,
25 const Convolution2dDescriptor& descriptor,
26 const TensorInfo& weights,
27 const Optional<TensorInfo>& biases,
28 bool isFastMathEnabled,
29 const ActivationDescriptor* activationDescriptor)
30 {
31 const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
32 const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
33 arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
34 aclWeightsInfo.set_are_values_constant(weights.IsConstant());
35
36 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX,
37 descriptor.m_DilationY);
38
39 arm_compute::TensorInfo aclBiasesInfo;
40 arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
41
42 if (descriptor.m_BiasEnabled)
43 {
44 ARMNN_ASSERT(biases.has_value());
45 // Same for bias as weights. We don't currently support non const.
46 if (!biases.value().IsConstant())
47 {
48 return arm_compute::Status{arm_compute::ErrorCode::RUNTIME_ERROR,
49 "ArmNN ClConvolution2dWorkload does not support non constant bias."};
50 }
51 aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
52 aclBiasesInfo.set_are_values_constant(biases.value().IsConstant());
53 optionalAclBiasesInfo = &aclBiasesInfo;
54 }
55
56 arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
57
58 const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
59 activationDescriptor);
60
61 return arm_compute::CLConvolutionLayer::validate(&aclInputInfo,
62 &aclWeightsInfo,
63 optionalAclBiasesInfo,
64 &aclOutputInfo,
65 layerInfo,
66 arm_compute::WeightsInfo(),
67 aclDilationInfo,
68 activationInfo,
69 isFastMathEnabled);
70 }
71
ClConvolution2dWorkload(const Convolution2dQueueDescriptor & descriptor,const WorkloadInfo & info,std::shared_ptr<arm_compute::MemoryManagerOnDemand> & memoryManager,const arm_compute::CLCompileContext & clCompileContext,const bool isFastMathEnabled)72 ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescriptor& descriptor,
73 const WorkloadInfo& info,
74 std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
75 const arm_compute::CLCompileContext& clCompileContext,
76 const bool isFastMathEnabled)
77 : ClBaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
78 , m_ConvolutionLayer(memoryManager)
79 {
80 ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution2dWorkload");
81
82 const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(m_Data.m_Parameters.m_DilationX,
83 m_Data.m_Parameters.m_DilationY);
84
85 uint32_t numInputs = m_Data.m_Parameters.m_BiasEnabled ? 3: 2;
86 m_Data.ValidateInputsOutputs("ClConvolution2dWorkload", numInputs, 1);
87
88 arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
89 arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
90 arm_compute::ICLTensor& weights = static_cast<IClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
91 if (m_Data.m_Parameters.m_BiasEnabled)
92 {
93 arm_compute::ICLTensor& bias = static_cast<IClTensorHandle*>(m_Data.m_Inputs[2])->GetTensor();
94 m_BiasProxy = std::make_unique<ICLTensorProxy>(&bias);
95 }
96
97 // Create Proxy tensor and set the initial tensor handle to it
98 m_InputProxy = std::make_unique<ICLTensorProxy>(&input);
99 m_OutputProxy = std::make_unique<ICLTensorProxy>(&output);
100 m_WeightsProxy = std::make_unique<ICLTensorProxy>(&weights);
101
102 arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
103 input.info()->set_data_layout(aclDataLayout);
104 output.info()->set_data_layout(aclDataLayout);
105 weights.info()->set_data_layout(aclDataLayout);
106
107 arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
108
109 const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
110
111 {
112 ARMNN_SCOPED_PROFILING_EVENT(Compute::Undefined, "ClConvolution2dWorkload_configure");
113 m_ConvolutionLayer.configure(clCompileContext,
114 m_InputProxy.get(),
115 m_WeightsProxy.get(),
116 m_BiasProxy.get(),
117 m_OutputProxy.get(),
118 padStrideInfo,
119 arm_compute::WeightsInfo(),
120 aclDilationInfo,
121 activationInfo,
122 isFastMathEnabled);
123 }
124
125 m_ConvolutionMethod =
126 m_ConvolutionLayer.get_convolution_method(input.info(),
127 weights.info(),
128 output.info(),
129 padStrideInfo,
130 arm_compute::WeightsInfo(),
131 activationInfo,
132 arm_compute::CLScheduler::get().target(),
133 aclDilationInfo,
134 isFastMathEnabled);
135
136 // Add details for profiling output
137 WorkloadInfo detailsInfo;
138
139 detailsInfo.m_InputTensorInfos = info.m_InputTensorInfos;
140 detailsInfo.m_OutputTensorInfos = info.m_OutputTensorInfos;
141 detailsInfo.m_WeightsTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[1]);
142 detailsInfo.m_ConvolutionMethod = armnn::Optional<std::string>(GetConvolutionMethodString(m_ConvolutionMethod));
143 if (descriptor.m_Parameters.m_BiasEnabled)
144 {
145 detailsInfo.m_BiasTensorInfo = armnn::Optional<armnn::TensorInfo>(info.m_InputTensorInfos[2]);
146 }
147
148 // Report Profiling Details
149 ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClConvolution2dWorkload_Construct",
150 descriptor.m_Parameters,
151 detailsInfo,
152 GetGuid());
153 }
154
Execute() const155 void ClConvolution2dWorkload::Execute() const
156 {
157 ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClConvolution2dWorkload_Execute", GetGuid());
158 RunClFunction(m_ConvolutionLayer, CHECK_LOCATION());
159 }
160
GetConvolutionMethod() const161 arm_compute::ConvolutionMethod ClConvolution2dWorkload::GetConvolutionMethod() const
162 {
163 return m_ConvolutionMethod;
164 }
165
Reconfigure()166 void ClConvolution2dWorkload::Reconfigure()
167 {
168 arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
169 arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
170
171 m_InputProxy->set(&input);
172 m_OutputProxy->set(&output);
173 }
174
175 } //namespace armnn
176