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
|
//
// Copyright © 2020 Arm Ltd and Contributors. All rights reserved.
// SPDX-License-Identifier: MIT
//
#pragma once
#include <armnn_delegate.hpp>
#include <armnnUtils/FloatingPointComparison.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> CreateSoftmaxTfLiteModel(tflite::BuiltinOperator softmaxOperatorCode,
tflite::TensorType tensorType,
const std::vector <int32_t>& tensorShape,
float beta)
{
using namespace tflite;
flatbuffers::FlatBufferBuilder flatBufferBuilder;
std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
buffers.push_back(CreateBuffer(flatBufferBuilder, flatBufferBuilder.CreateVector({})));
std::array<flatbuffers::Offset<Tensor>, 2> tensors;
tensors[0] = CreateTensor(flatBufferBuilder,
flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(),
tensorShape.size()),
tensorType,
0);
tensors[1] = CreateTensor(flatBufferBuilder,
flatBufferBuilder.CreateVector<int32_t>(tensorShape.data(),
tensorShape.size()),
tensorType,
0);
const std::vector<int32_t> operatorInputs({0});
const std::vector<int32_t> operatorOutputs({1});
flatbuffers::Offset<Operator> softmaxOperator;
flatbuffers::Offset<flatbuffers::String> modelDescription;
flatbuffers::Offset<OperatorCode> operatorCode;
switch (softmaxOperatorCode)
{
case tflite::BuiltinOperator_SOFTMAX:
softmaxOperator =
CreateOperator(flatBufferBuilder,
0,
flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
BuiltinOptions_SoftmaxOptions,
CreateSoftmaxOptions(flatBufferBuilder, beta).Union());
modelDescription = flatBufferBuilder.CreateString("ArmnnDelegate: Softmax Operator Model");
operatorCode = CreateOperatorCode(flatBufferBuilder,
tflite::BuiltinOperator_SOFTMAX);
break;
case tflite::BuiltinOperator_LOG_SOFTMAX:
softmaxOperator =
CreateOperator(flatBufferBuilder,
0,
flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(), operatorOutputs.size()),
BuiltinOptions_LogSoftmaxOptions,
CreateLogSoftmaxOptions(flatBufferBuilder).Union());
flatBufferBuilder.CreateString("ArmnnDelegate: Log-Softmax Operator Model");
operatorCode = CreateOperatorCode(flatBufferBuilder,
tflite::BuiltinOperator_LOG_SOFTMAX);
break;
default:
break;
}
const std::vector<int32_t> subgraphInputs({0});
const std::vector<int32_t> subgraphOutputs({1});
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(&softmaxOperator, 1));
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());
}
void SoftmaxTest(tflite::BuiltinOperator softmaxOperatorCode,
tflite::TensorType tensorType,
std::vector<armnn::BackendId>& backends,
std::vector<int32_t>& shape,
std::vector<float>& inputValues,
std::vector<float>& expectedOutputValues,
float beta = 0)
{
using namespace tflite;
std::vector<char> modelBuffer = CreateSoftmaxTfLiteModel(softmaxOperatorCode,
tensorType,
shape,
beta);
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
auto tfLiteDelegateInputId = tfLiteInterpreter->inputs()[0];
auto tfLiteInterpreterInputData = tfLiteInterpreter->typed_tensor<float>(tfLiteDelegateInputId);
for (unsigned int i = 0; i < inputValues.size(); ++i)
{
tfLiteInterpreterInputData[i] = inputValues[i];
}
auto armnnDelegateInputId = armnnDelegateInterpreter->inputs()[0];
auto armnnDelegateInputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateInputId);
for (unsigned int i = 0; i < inputValues.size(); ++i)
{
armnnDelegateInputData[i] = inputValues[i];
}
// Run EnqueWorkload
CHECK(tfLiteInterpreter->Invoke() == kTfLiteOk);
CHECK(armnnDelegateInterpreter->Invoke() == kTfLiteOk);
// Compare output data
auto tfLiteInterpreterOutputId = tfLiteInterpreter->outputs()[0];
auto tfLiteInterpreterOutputData = tfLiteInterpreter->typed_tensor<float>(tfLiteInterpreterOutputId);
auto armnnDelegateOutputId = armnnDelegateInterpreter->outputs()[0];
auto armnnDelegateOutputData = armnnDelegateInterpreter->typed_tensor<float>(armnnDelegateOutputId);
for (size_t i = 0; i < inputValues.size(); ++i)
{
CHECK(armnnUtils::within_percentage_tolerance(expectedOutputValues[i], armnnDelegateOutputData[i], 1e-5));
CHECK(armnnUtils::within_percentage_tolerance(tfLiteInterpreterOutputData[i],
armnnDelegateOutputData[i], 1e-5));
}
}
} // anonymous namespace
|
