1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
|
//
// Copyright © 2017 Arm Ltd. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ClConvolution2dWorkload.hpp"
#include "ClWorkloadUtils.hpp"
#include <cl/ClLayerSupport.hpp>
#include <cl/ClTensorHandle.hpp>
#include <cl/ClLayerSupport.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <backendsCommon/TensorHandle.hpp>
#include <arm_compute/runtime/CL/functions/CLConvolutionLayer.h>
namespace armnn
{
using namespace armcomputetensorutils;
arm_compute::Status ClConvolution2dWorkloadValidate(const TensorInfo& input,
const TensorInfo& output,
const Convolution2dDescriptor& descriptor,
const TensorInfo& weights,
const Optional<TensorInfo>& biases,
bool isFastMathEnabled,
const ActivationDescriptor* activationDescriptor)
{
const arm_compute::TensorInfo aclInputInfo = BuildArmComputeTensorInfo(input, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclOutputInfo = BuildArmComputeTensorInfo(output, descriptor.m_DataLayout);
const arm_compute::TensorInfo aclWeightsInfo = BuildArmComputeTensorInfo(weights, descriptor.m_DataLayout);
const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(descriptor.m_DilationX,
descriptor.m_DilationY);
arm_compute::TensorInfo aclBiasesInfo;
arm_compute::TensorInfo *optionalAclBiasesInfo = nullptr;
if (descriptor.m_BiasEnabled)
{
ARMNN_ASSERT(biases.has_value());
aclBiasesInfo = BuildArmComputeTensorInfo(biases.value(), descriptor.m_DataLayout);
optionalAclBiasesInfo = &aclBiasesInfo;
}
arm_compute::PadStrideInfo layerInfo = BuildArmComputePadStrideInfo(descriptor);
const arm_compute::ActivationLayerInfo activationInfo = ConvertActivationDescriptorToAclActivationLayerInfo(
activationDescriptor);
return arm_compute::CLConvolutionLayer::validate(&aclInputInfo,
&aclWeightsInfo,
optionalAclBiasesInfo,
&aclOutputInfo,
layerInfo,
arm_compute::WeightsInfo(),
aclDilationInfo,
activationInfo,
isFastMathEnabled);
}
ClConvolution2dWorkload::ClConvolution2dWorkload(const Convolution2dQueueDescriptor& descriptor,
const WorkloadInfo& info,
std::shared_ptr<arm_compute::MemoryManagerOnDemand>& memoryManager,
const arm_compute::CLCompileContext& clCompileContext,
const bool isFastMathEnabled)
: BaseWorkload<Convolution2dQueueDescriptor>(descriptor, info)
, m_ConvolutionLayer(memoryManager)
{
// todo: check tensor shapes match.
const TensorInfo& weightInfo = m_Data.m_Weight->GetTensorInfo();
m_KernelTensor = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_KernelTensor, weightInfo, m_Data.m_Parameters.m_DataLayout);
const arm_compute::Size2D aclDilationInfo = BuildArmComputeSize2D(m_Data.m_Parameters.m_DilationX,
m_Data.m_Parameters.m_DilationY);
if (m_Data.m_Parameters.m_BiasEnabled)
{
m_BiasTensor = std::make_unique<arm_compute::CLTensor>();
BuildArmComputeTensor(*m_BiasTensor, m_Data.m_Bias->GetTensorInfo(), m_Data.m_Parameters.m_DataLayout);
}
m_Data.ValidateInputsOutputs("ClConvolution2dWorkload", 1, 1);
arm_compute::ICLTensor& input = static_cast<IClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
arm_compute::ICLTensor& output = static_cast<IClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
arm_compute::DataLayout aclDataLayout = ConvertDataLayout(m_Data.m_Parameters.m_DataLayout);
input.info()->set_data_layout(aclDataLayout);
output.info()->set_data_layout(aclDataLayout);
arm_compute::PadStrideInfo padStrideInfo = BuildArmComputePadStrideInfo(m_Data.m_Parameters);
const arm_compute::ActivationLayerInfo activationInfo = ConvertAdditionalInfoToAclActivationLayerInfo(descriptor);
m_ConvolutionLayer.configure(clCompileContext,
&input,
m_KernelTensor.get(),
m_BiasTensor.get(),
&output,
padStrideInfo,
arm_compute::WeightsInfo(),
aclDilationInfo,
activationInfo,
isFastMathEnabled);
m_ConvolutionMethod =
m_ConvolutionLayer.get_convolution_method(input.info(),
m_KernelTensor->info(),
output.info(),
padStrideInfo,
arm_compute::WeightsInfo(),
activationInfo,
arm_compute::CLScheduler::get().target(),
aclDilationInfo,
isFastMathEnabled);
InitializeArmComputeClTensorData(*m_KernelTensor, m_Data.m_Weight);
if (m_BiasTensor)
{
InitializeArmComputeClTensorData(*m_BiasTensor, m_Data.m_Bias);
}
// Force Compute Library to perform the necessary copying and reshaping, after which
// delete all the input tensors that will no longer be needed
m_ConvolutionLayer.prepare();
FreeUnusedTensors();
}
void ClConvolution2dWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_CL("ClConvolution2dWorkload_Execute");
RunClFunction(m_ConvolutionLayer, CHECK_LOCATION());
}
arm_compute::ConvolutionMethod ClConvolution2dWorkload::GetConvolutionMethod() const
{
return m_ConvolutionMethod;
}
void ClConvolution2dWorkload::FreeUnusedTensors()
{
FreeTensorIfUnused(m_KernelTensor);
FreeTensorIfUnused(m_BiasTensor);
}
} //namespace armnn
|
