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
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
|
//
// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#include "ClBatchMatMulWorkload.hpp"
#include "ClWorkloadUtils.hpp"
#include <aclCommon/ArmComputeTensorUtils.hpp>
#include <aclCommon/ArmComputeUtils.hpp>
#include <armnn/utility/PolymorphicDowncast.hpp>
#include <armnnUtils/Permute.hpp>
#include <armnnUtils/TensorUtils.hpp>
#include <backendsCommon/WorkloadUtils.hpp>
#include <cl/ClTensorHandle.hpp>
#include <arm_compute/runtime/CL/functions/CLGEMM.h>
#include <arm_compute/runtime/CL/functions/CLPermute.h>
namespace armnn
{
arm_compute::Status ClBatchMatMulValidate(const TensorInfo& inputX,
const TensorInfo& inputY,
const TensorInfo& output,
const BatchMatMulDescriptor& descriptor)
{
if (descriptor.m_AdjointX || descriptor.m_AdjointY )
{
throw Exception("Support for adjoint not implemented.");
}
if (descriptor.m_DataLayoutX != armnn::DataLayout::NCHW || descriptor.m_DataLayoutY != armnn::DataLayout::NCHW )
{
throw Exception("Only supported the MatMul in the last 2 dimensions");
}
arm_compute::Status statusGEMM = arm_compute::Status(arm_compute::ErrorCode::OK);
arm_compute::Status statusPermuteX = arm_compute::Status(arm_compute::ErrorCode::OK);
arm_compute::Status statusPermuteY = arm_compute::Status(arm_compute::ErrorCode::OK);
// ClGemmMatrixMultiplyNativeKernel used by CLGEMM can only support 3 dimensional
// tensors so try to reduce the dimensions to 3
const auto aclInputXInfo = armcomputetensorutils::BuildArmComputeTensorInfo(inputX, descriptor.m_DataLayoutX, 3);
const auto aclInputYInfo = armcomputetensorutils::BuildArmComputeTensorInfo(inputY, descriptor.m_DataLayoutY, 3);
const auto aclOutputInfo = armcomputetensorutils::BuildArmComputeTensorInfo(output, descriptor.m_DataLayoutY, 3);
arm_compute::TensorInfo aclPermutedXInfo = arm_compute::TensorInfo();
arm_compute::TensorInfo aclPermutedYInfo = arm_compute::TensorInfo();
if (descriptor.m_TransposeX == true)
{
armnn::TensorInfo inputXStripped = armnnUtils::ReduceDims(inputX, 3);
auto permutationXVector = GeneratePermutationVectorOnLastTwoDimensions(inputXStripped.GetNumDimensions());
const auto aclPermutationXVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationXVector);
const TensorInfo permutedXInfo = armnnUtils::Permuted(inputXStripped, permutationXVector);
aclPermutedXInfo = armcomputetensorutils::BuildArmComputeTensorInfo(permutedXInfo, 3);
statusPermuteX = arm_compute::CLPermute::validate(&aclInputXInfo,
&aclPermutedXInfo,
aclPermutationXVector);
}
if (descriptor.m_TransposeY == true)
{
armnn::TensorInfo inputYStripped = armnnUtils::ReduceDims(inputY, 3);
auto permutationYVector = GeneratePermutationVectorOnLastTwoDimensions(inputYStripped.GetNumDimensions());
const auto aclPermutationYVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationYVector);
const TensorInfo permutedYInfo = armnnUtils::Permuted(inputYStripped, permutationYVector);
aclPermutedYInfo = armcomputetensorutils::BuildArmComputeTensorInfo(permutedYInfo, 3);
statusPermuteY = arm_compute::CLPermute::validate(&aclInputYInfo,
&aclPermutedYInfo,
aclPermutationYVector);
}
const arm_compute::GEMMInfo& gemm_info = arm_compute::GEMMInfo(false, // is inputX reshaped
false, // is inputY reshaped
false); // is inputY reshaped only 1st run
statusGEMM = arm_compute::CLGEMM::validate(descriptor.m_TransposeX ? &aclPermutedXInfo : &aclInputXInfo,
descriptor.m_TransposeY ? &aclPermutedYInfo : &aclInputYInfo,
nullptr,
&aclOutputInfo,
1.0,
0,
gemm_info);
if (statusPermuteX.error_code() == arm_compute::ErrorCode::OK &&
statusPermuteY.error_code() == arm_compute::ErrorCode::OK &&
statusGEMM.error_code() == arm_compute::ErrorCode::OK)
{
return arm_compute::Status(arm_compute::ErrorCode::OK,
"All Batch Mat Mul layers validate status OK.");
}
else
{
return arm_compute::Status(arm_compute::ErrorCode::RUNTIME_ERROR,
"BatchMatMul layer validate status failed."
+ statusGEMM.error_description()
+ statusPermuteX.error_description()
+ statusPermuteY.error_description());
}
}
ClBatchMatMulWorkload::ClBatchMatMulWorkload(const BatchMatMulQueueDescriptor& descriptor,
const WorkloadInfo& info,
const arm_compute::CLCompileContext& clCompileContext)
: ClBaseWorkload<BatchMatMulQueueDescriptor>(descriptor, info)
{
// Report Profiling Details
ARMNN_REPORT_PROFILING_WORKLOAD_DESC("ClBatchMatMulWorkload_Construct",
descriptor.m_Parameters,
info,
this->GetGuid());
if (descriptor.m_Parameters.m_AdjointX || descriptor.m_Parameters.m_AdjointY )
{
throw Exception("Support for adjoint not implemented.");
}
if (descriptor.m_Parameters.m_DataLayoutX != armnn::DataLayout::NCHW ||
descriptor.m_Parameters.m_DataLayoutY != armnn::DataLayout::NCHW )
{
throw Exception("Only supported the MatMul in the last 2 dimensions");
}
m_Data.ValidateInputsOutputs("ClBatchMatMulWorkload", 2, 1);
const arm_compute::ICLTensor& inputX = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[0])->GetTensor();
const arm_compute::ICLTensor& inputY = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Inputs[1])->GetTensor();
arm_compute::ICLTensor& output = PolymorphicDowncast<ClTensorHandle*>(m_Data.m_Outputs[0])->GetTensor();
inputX.info()->set_data_layout(armcomputetensorutils::ConvertDataLayout(m_Data.m_Parameters.m_DataLayoutX));
arm_compute::TensorShape inputXTensorInfo = armcomputetensorutils::BuildArmComputeTensorShape(
info.m_InputTensorInfos[0].GetShape(), 3);
inputX.info()->set_tensor_shape(inputXTensorInfo);
inputY.info()->set_data_layout(armcomputetensorutils::ConvertDataLayout(m_Data.m_Parameters.m_DataLayoutY));
arm_compute::TensorShape inputYTensorInfo = armcomputetensorutils::BuildArmComputeTensorShape(
info.m_InputTensorInfos[1].GetShape(), 3);
inputY.info()->set_tensor_shape(inputYTensorInfo);
arm_compute::TensorInfo aclPermutedXInfo = arm_compute::TensorInfo();
arm_compute::TensorInfo aclPermutedYInfo = arm_compute::TensorInfo();
if (descriptor.m_Parameters.m_TransposeX == true)
{
armnn::TensorInfo strippedInfo = armnnUtils::ReduceDims(info.m_InputTensorInfos[0], 3);
armnn::PermutationVector permutationXVector
= GeneratePermutationVectorOnLastTwoDimensions(strippedInfo.GetNumDimensions());
const TensorInfo permutedXInfo = armnnUtils::Permuted(strippedInfo, permutationXVector);
const auto aclPermutationXVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationXVector);
armcomputetensorutils::BuildArmComputeTensor(m_PermutedTensorX, permutedXInfo);
armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermutedTensorX);
auto permuteLayerX = std::make_unique<arm_compute::CLPermute>();
permuteLayerX->configure(clCompileContext,
&inputX,
&m_PermutedTensorX,
aclPermutationXVector);
m_PermuteLayerX.reset(permuteLayerX.release());
}
if (descriptor.m_Parameters.m_TransposeY == true)
{
armnn::TensorInfo strippedInfo = armnnUtils::ReduceDims(info.m_InputTensorInfos[1], 3);
armnn::PermutationVector permutationYVector
= GeneratePermutationVectorOnLastTwoDimensions(strippedInfo.GetNumDimensions());
const TensorInfo permutedYInfo = armnnUtils::Permuted(strippedInfo, permutationYVector);
const auto aclPermutationYVector = armcomputetensorutils::BuildArmComputePermutationVector(permutationYVector);
armcomputetensorutils::BuildArmComputeTensor(m_PermutedTensorY, permutedYInfo);
armcomputetensorutils::InitialiseArmComputeTensorEmpty(m_PermutedTensorY);
auto permuteLayerY = std::make_unique<arm_compute::CLPermute>();
permuteLayerY->configure(clCompileContext,
&inputY,
&m_PermutedTensorY,
aclPermutationYVector);
m_PermuteLayerY.reset(permuteLayerY.release());
}
const arm_compute::GEMMInfo& gemm_info = arm_compute::GEMMInfo(false, // is inputX reshaped
false, // is inputY reshaped
false); // is inputY reshaped only 1st run
auto gemmLayer = std::make_unique<arm_compute::CLGEMM>();
gemmLayer->configure(clCompileContext,
descriptor.m_Parameters.m_TransposeX ? &m_PermutedTensorX : &inputX,
descriptor.m_Parameters.m_TransposeY ? &m_PermutedTensorY : &inputY,
nullptr,
&output,
1.0,
0,
gemm_info);
m_GEMMLayer.reset(gemmLayer.release());
}
void ClBatchMatMulWorkload::Execute() const
{
ARMNN_SCOPED_PROFILING_EVENT_CL_GUID("ClBatchMatMulWorkload_Execute", this->GetGuid());
if (m_PermuteLayerX)
{
m_PermuteLayerX->run();
}
if (m_PermuteLayerY)
{
m_PermuteLayerY->run();
}
m_GEMMLayer->run();
}
} //namespace armnn
|
