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
|
//
// Copyright © 2021 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>
#include <string>
namespace
{
std::vector<char> CreateReduceTfLiteModel(tflite::BuiltinOperator reduceOperatorCode,
tflite::TensorType tensorType,
std::vector<int32_t>& input0TensorShape,
std::vector<int32_t>& input1TensorShape,
const std::vector <int32_t>& outputTensorShape,
std::vector<int32_t>& axisData,
const bool keepDims,
float quantScale = 1.0f,
int quantOffset = 0)
{
using namespace tflite;
flatbuffers::FlatBufferBuilder flatBufferBuilder;
std::array<flatbuffers::Offset<tflite::Buffer>, 2> buffers;
buffers[0] = CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({}));
buffers[1] = CreateBuffer(flatBufferBuilder,
flatBufferBuilder.CreateVector(reinterpret_cast<const uint8_t*>(axisData.data()),
sizeof(int32_t) * axisData.size()));
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>(input0TensorShape.data(),
input0TensorShape.size()),
tensorType,
0,
flatBufferBuilder.CreateString("input"),
quantizationParameters);
tensors[1] = CreateTensor(flatBufferBuilder,
flatBufferBuilder.CreateVector<int32_t>(input1TensorShape.data(),
input1TensorShape.size()),
::tflite::TensorType_INT32,
1,
flatBufferBuilder.CreateString("axis"),
quantizationParameters);
// Create output tensor
tensors[2] = CreateTensor(flatBufferBuilder,
flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(),
outputTensorShape.size()),
tensorType,
0,
flatBufferBuilder.CreateString("output"),
quantizationParameters);
// Create operator. Reduce operations MIN, MAX, SUM, MEAN uses ReducerOptions.
tflite::BuiltinOptions operatorBuiltinOptionsType = tflite::BuiltinOptions_ReducerOptions;
flatbuffers::Offset<void> operatorBuiltinOptions = CreateReducerOptions(flatBufferBuilder, keepDims).Union();
const std::vector<int> operatorInputs{ {0, 1} };
const std::vector<int> operatorOutputs{ 2 };
flatbuffers::Offset <Operator> reduceOperator =
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(&reduceOperator, 1));
flatbuffers::Offset <flatbuffers::String> modelDescription =
flatBufferBuilder.CreateString("ArmnnDelegate: Reduce Operator Model");
flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, reduceOperatorCode);
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 ReduceTest(tflite::BuiltinOperator reduceOperatorCode,
tflite::TensorType tensorType,
std::vector<armnn::BackendId>& backends,
std::vector<int32_t>& input0Shape,
std::vector<int32_t>& input1Shape,
std::vector<int32_t>& expectedOutputShape,
std::vector<T>& input0Values,
std::vector<int32_t>& input1Values,
std::vector<T>& expectedOutputValues,
const bool keepDims,
float quantScale = 1.0f,
int quantOffset = 0)
{
using namespace tflite;
std::vector<char> modelBuffer = CreateReduceTfLiteModel(reduceOperatorCode,
tensorType,
input0Shape,
input1Shape,
expectedOutputShape,
input1Values,
keepDims,
quantScale,
quantOffset);
const Model* tfLiteModel = GetModel(modelBuffer.data());
// 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
armnnDelegate::FillInput<T>(tfLiteInterpreter, 0, input0Values);
armnnDelegate::FillInput<T>(armnnDelegateInterpreter, 0, input0Values);
// Run EnqueWorkload
CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
// Compare output data
armnnDelegate::CompareOutputData<T>(tfLiteInterpreter,
armnnDelegateInterpreter,
expectedOutputShape,
expectedOutputValues);
armnnDelegateInterpreter.reset(nullptr);
}
} // anonymous namespace
|
