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
|
//
// Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include "TestUtils.hpp"
#include <armnn_delegate.hpp>
#include <flatbuffers/flatbuffers.h>
#include <tensorflow/lite/interpreter.h>
#include <tensorflow/lite/kernels/register.h>
#include <tensorflow/lite/model.h>
#include <tensorflow/lite/schema/schema_generated.h>
#include <tensorflow/lite/version.h>
#include <doctest/doctest.h>
namespace
{
std::vector<char> CreateBatchMatMulTfLiteModel(
tflite::BuiltinOperator bmmOperatorCode,
tflite::TensorType tensorType,
const std::vector <int32_t>& LHSInputTensorShape,
const std::vector <int32_t>& RHSInputTensorShape,
const std::vector <int32_t>& outputTensorShape,
bool adjX = false,
bool adjY = false,
float quantScale = 1.0f,
int quantOffset = 0)
{
using namespace tflite;
flatbuffers::FlatBufferBuilder flatBufferBuilder;
std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
buffers.push_back(CreateBuffer(flatBufferBuilder));
buffers.push_back(CreateBuffer(flatBufferBuilder));
buffers.push_back(CreateBuffer(flatBufferBuilder));
buffers.push_back(CreateBuffer(flatBufferBuilder));
auto quantizationParameters =
CreateQuantizationParameters(flatBufferBuilder,
0,
0,
flatBufferBuilder.CreateVector<float>({ quantScale }),
flatBufferBuilder.CreateVector<int64_t>({ quantOffset }));
std::array<flatbuffers::Offset<Tensor>, 3> tensors;
tensors[0] = CreateTensor(flatBufferBuilder,
flatBufferBuilder.CreateVector<int32_t>(LHSInputTensorShape.data(),
LHSInputTensorShape.size()),
tensorType,
1,
flatBufferBuilder.CreateString("LHSInput"),
quantizationParameters);
tensors[1] = CreateTensor(flatBufferBuilder,
flatBufferBuilder.CreateVector<int32_t>(RHSInputTensorShape.data(),
RHSInputTensorShape.size()),
tensorType,
2,
flatBufferBuilder.CreateString("RHSInput"),
quantizationParameters);
tensors[2] = CreateTensor(flatBufferBuilder,
flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(),
outputTensorShape.size()),
tensorType,
3,
flatBufferBuilder.CreateString("output"),
quantizationParameters);
// create operator
tflite::BuiltinOptions operatorBuiltinOptionsType = BuiltinOptions_BatchMatMulOptions;
flatbuffers::Offset<void> operatorBuiltinOptions = CreateBatchMatMulOptions(flatBufferBuilder,
adjX,
adjY).Union();
const std::vector<int32_t> operatorInputs{{0, 1}};
const std::vector<int32_t> operatorOutputs{2};
flatbuffers::Offset <Operator> bmmOperator =
CreateOperator(flatBufferBuilder,
0,
flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(),
operatorOutputs.size()),
operatorBuiltinOptionsType,
operatorBuiltinOptions);
const std::vector<int> subgraphInputs{{0, 1}};
const std::vector<int> subgraphOutputs{2};
flatbuffers::Offset <SubGraph> subgraph =
CreateSubGraph(flatBufferBuilder,
flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()),
flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(),
subgraphOutputs.size()),
flatBufferBuilder.CreateVector(&bmmOperator, 1));
flatbuffers::Offset <flatbuffers::String> modelDescription =
flatBufferBuilder.CreateString("ArmnnDelegate: BatchMatMul Operator Model");
flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, bmmOperatorCode);
flatbuffers::Offset <Model> flatbufferModel =
CreateModel(flatBufferBuilder,
TFLITE_SCHEMA_VERSION,
flatBufferBuilder.CreateVector(&operatorCode, 1),
flatBufferBuilder.CreateVector(&subgraph, 1),
modelDescription,
flatBufferBuilder.CreateVector(buffers.data(), buffers.size()));
flatBufferBuilder.Finish(flatbufferModel);
return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
}
template <typename T>
void BatchMatMulTest(tflite::BuiltinOperator bmmOperatorCode,
tflite::TensorType tensorType,
std::vector<armnn::BackendId>& backends,
std::vector<int32_t>& LHSInputShape,
std::vector<int32_t>& RHSInputShape,
std::vector<int32_t>& outputShape,
std::vector<T>& LHSInputValues,
std::vector<T>& RHSInputValues,
std::vector<T>& expectedOutputValues,
bool adjX = false,
bool adjY = false,
float quantScale = 1.0f,
int quantOffset = 0)
{
using namespace tflite;
std::vector<char> modelBuffer = CreateBatchMatMulTfLiteModel(bmmOperatorCode,
tensorType,
LHSInputShape,
RHSInputShape,
outputShape,
adjX,
adjY,
quantScale,
quantOffset);
const Model* tfLiteModel = GetModel(modelBuffer.data());
CHECK(tfLiteModel != nullptr);
// Create TfLite Interpreters
std::unique_ptr<Interpreter> armnnDelegateInterpreter;
CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
(&armnnDelegateInterpreter) == kTfLiteOk);
CHECK(armnnDelegateInterpreter != nullptr);
CHECK(armnnDelegateInterpreter->AllocateTensors() == kTfLiteOk);
std::unique_ptr<Interpreter> tfLiteInterpreter;
CHECK(InterpreterBuilder(tfLiteModel, ::tflite::ops::builtin::BuiltinOpResolver())
(&tfLiteInterpreter) == kTfLiteOk);
CHECK(tfLiteInterpreter != nullptr);
CHECK(tfLiteInterpreter->AllocateTensors() == kTfLiteOk);
// Create the ArmNN Delegate
armnnDelegate::DelegateOptions delegateOptions(backends);
std::unique_ptr<TfLiteDelegate, decltype(&armnnDelegate::TfLiteArmnnDelegateDelete)>
theArmnnDelegate(armnnDelegate::TfLiteArmnnDelegateCreate(delegateOptions),
armnnDelegate::TfLiteArmnnDelegateDelete);
CHECK(theArmnnDelegate != nullptr);
// Modify armnnDelegateInterpreter to use armnnDelegate
CHECK(armnnDelegateInterpreter->ModifyGraphWithDelegate(theArmnnDelegate.get()) == kTfLiteOk);
// Set input data
auto tfLiteDelegateLHSInputId = tfLiteInterpreter->inputs()[0];
auto tfLiteDelegateLHSInputData = tfLiteInterpreter->typed_tensor<T>(tfLiteDelegateLHSInputId);
auto tfLiteDelegateRHSInputId = tfLiteInterpreter->inputs()[1];
auto tfLiteDelegateRHSInputData = tfLiteInterpreter->typed_tensor<T>(tfLiteDelegateRHSInputId);
for (unsigned int i = 0; i < LHSInputValues.size(); ++i)
{
tfLiteDelegateLHSInputData[i] = LHSInputValues[i];
}
for (unsigned int i = 0; i < RHSInputValues.size(); ++i)
{
tfLiteDelegateRHSInputData[i] = RHSInputValues[i];
}
auto armnnDelegateLHSInputId = armnnDelegateInterpreter->inputs()[0];
auto armnnDelegateLHSInputData = armnnDelegateInterpreter->typed_tensor<T>(armnnDelegateLHSInputId);
auto armnnDelegateRHSInputId = armnnDelegateInterpreter->inputs()[1];
auto armnnDelegateRHSInputData = armnnDelegateInterpreter->typed_tensor<T>(armnnDelegateRHSInputId);
for (unsigned int i = 0; i < LHSInputValues.size(); ++i)
{
armnnDelegateLHSInputData[i] = LHSInputValues[i];
}
for (unsigned int i = 0; i < RHSInputValues.size(); ++i)
{
armnnDelegateRHSInputData[i] = RHSInputValues[i];
}
// Run EnqueueWorkload
CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
armnnDelegate::CompareOutputData(tfLiteInterpreter, armnnDelegateInterpreter,
outputShape, expectedOutputValues);
}
} // anonymous namespace
|
