1 //
2 // Copyright © 2022-2023 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5
6 #pragma once
7
8 #include "TestUtils.hpp"
9
10 #include <armnn_delegate.hpp>
11 #include <DelegateTestInterpreter.hpp>
12
13 #include <flatbuffers/flatbuffers.h>
14 #include <tensorflow/lite/kernels/register.h>
15 #include <tensorflow/lite/version.h>
16
17 #include <schema_generated.h>
18
19 #include <doctest/doctest.h>
20
21 namespace
22 {
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)23 std::vector<char> CreateBatchMatMulTfLiteModel(
24 tflite::BuiltinOperator bmmOperatorCode,
25 tflite::TensorType tensorType,
26 const std::vector <int32_t>& LHSInputTensorShape,
27 const std::vector <int32_t>& RHSInputTensorShape,
28 const std::vector <int32_t>& outputTensorShape,
29 bool adjX = false,
30 bool adjY = false,
31 float quantScale = 1.0f,
32 int quantOffset = 0)
33 {
34 using namespace tflite;
35 flatbuffers::FlatBufferBuilder flatBufferBuilder;
36
37 std::vector<flatbuffers::Offset<tflite::Buffer>> buffers;
38 buffers.push_back(CreateBuffer(flatBufferBuilder));
39 buffers.push_back(CreateBuffer(flatBufferBuilder));
40 buffers.push_back(CreateBuffer(flatBufferBuilder));
41 buffers.push_back(CreateBuffer(flatBufferBuilder));
42
43 auto quantizationParameters =
44 CreateQuantizationParameters(flatBufferBuilder,
45 0,
46 0,
47 flatBufferBuilder.CreateVector<float>({ quantScale }),
48 flatBufferBuilder.CreateVector<int64_t>({ quantOffset }));
49
50 std::array<flatbuffers::Offset<Tensor>, 3> tensors;
51 tensors[0] = CreateTensor(flatBufferBuilder,
52 flatBufferBuilder.CreateVector<int32_t>(LHSInputTensorShape.data(),
53 LHSInputTensorShape.size()),
54 tensorType,
55 1,
56 flatBufferBuilder.CreateString("LHSInput"),
57 quantizationParameters);
58
59 tensors[1] = CreateTensor(flatBufferBuilder,
60 flatBufferBuilder.CreateVector<int32_t>(RHSInputTensorShape.data(),
61 RHSInputTensorShape.size()),
62 tensorType,
63 2,
64 flatBufferBuilder.CreateString("RHSInput"),
65 quantizationParameters);
66
67 tensors[2] = CreateTensor(flatBufferBuilder,
68 flatBufferBuilder.CreateVector<int32_t>(outputTensorShape.data(),
69 outputTensorShape.size()),
70 tensorType,
71 3,
72 flatBufferBuilder.CreateString("output"),
73 quantizationParameters);
74
75 // create operator
76 tflite::BuiltinOptions operatorBuiltinOptionsType = BuiltinOptions_BatchMatMulOptions;
77 flatbuffers::Offset<void> operatorBuiltinOptions = CreateBatchMatMulOptions(flatBufferBuilder,
78 adjX,
79 adjY).Union();
80
81 const std::vector<int32_t> operatorInputs{{0, 1}};
82 const std::vector<int32_t> operatorOutputs{2};
83 flatbuffers::Offset <Operator> bmmOperator =
84 CreateOperator(flatBufferBuilder,
85 0,
86 flatBufferBuilder.CreateVector<int32_t>(operatorInputs.data(), operatorInputs.size()),
87 flatBufferBuilder.CreateVector<int32_t>(operatorOutputs.data(),
88 operatorOutputs.size()),
89 operatorBuiltinOptionsType,
90 operatorBuiltinOptions);
91
92 const std::vector<int> subgraphInputs{{0, 1}};
93 const std::vector<int> subgraphOutputs{2};
94 flatbuffers::Offset <SubGraph> subgraph =
95 CreateSubGraph(flatBufferBuilder,
96 flatBufferBuilder.CreateVector(tensors.data(), tensors.size()),
97 flatBufferBuilder.CreateVector<int32_t>(subgraphInputs.data(), subgraphInputs.size()),
98 flatBufferBuilder.CreateVector<int32_t>(subgraphOutputs.data(),
99 subgraphOutputs.size()),
100 flatBufferBuilder.CreateVector(&bmmOperator, 1));
101
102 flatbuffers::Offset <flatbuffers::String> modelDescription =
103 flatBufferBuilder.CreateString("ArmnnDelegate: BatchMatMul Operator Model");
104 flatbuffers::Offset <OperatorCode> operatorCode = CreateOperatorCode(flatBufferBuilder, bmmOperatorCode);
105
106 flatbuffers::Offset <Model> flatbufferModel =
107 CreateModel(flatBufferBuilder,
108 TFLITE_SCHEMA_VERSION,
109 flatBufferBuilder.CreateVector(&operatorCode, 1),
110 flatBufferBuilder.CreateVector(&subgraph, 1),
111 modelDescription,
112 flatBufferBuilder.CreateVector(buffers.data(), buffers.size()));
113
114 flatBufferBuilder.Finish(flatbufferModel, armnnDelegate::FILE_IDENTIFIER);
115
116 return std::vector<char>(flatBufferBuilder.GetBufferPointer(),
117 flatBufferBuilder.GetBufferPointer() + flatBufferBuilder.GetSize());
118 }
119
120 template <typename T>
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)121 void BatchMatMulTest(tflite::BuiltinOperator bmmOperatorCode,
122 tflite::TensorType tensorType,
123 std::vector<armnn::BackendId>& backends,
124 std::vector<int32_t>& LHSInputShape,
125 std::vector<int32_t>& RHSInputShape,
126 std::vector<int32_t>& outputShape,
127 std::vector<T>& LHSInputValues,
128 std::vector<T>& RHSInputValues,
129 std::vector<T>& expectedOutputValues,
130 bool adjX = false,
131 bool adjY = false,
132 float quantScale = 1.0f,
133 int quantOffset = 0)
134 {
135 using namespace delegateTestInterpreter;
136 std::vector<char> modelBuffer = CreateBatchMatMulTfLiteModel(bmmOperatorCode,
137 tensorType,
138 LHSInputShape,
139 RHSInputShape,
140 outputShape,
141 adjX,
142 adjY,
143 quantScale,
144 quantOffset);
145
146 // Setup interpreter with just TFLite Runtime.
147 auto tfLiteInterpreter = DelegateTestInterpreter(modelBuffer);
148 CHECK(tfLiteInterpreter.AllocateTensors() == kTfLiteOk);
149 CHECK(tfLiteInterpreter.FillInputTensor<T>(LHSInputValues, 0) == kTfLiteOk);
150 CHECK(tfLiteInterpreter.FillInputTensor<T>(RHSInputValues, 1) == kTfLiteOk);
151 CHECK(tfLiteInterpreter.Invoke() == kTfLiteOk);
152 std::vector<T> tfLiteOutputValues = tfLiteInterpreter.GetOutputResult<T>(0);
153 std::vector<int32_t> tfLiteOutputShape = tfLiteInterpreter.GetOutputShape(0);
154
155 // Setup interpreter with Arm NN Delegate applied.
156 auto armnnInterpreter = DelegateTestInterpreter(modelBuffer, backends);
157 CHECK(armnnInterpreter.AllocateTensors() == kTfLiteOk);
158 CHECK(armnnInterpreter.FillInputTensor<T>(LHSInputValues, 0) == kTfLiteOk);
159 CHECK(armnnInterpreter.FillInputTensor<T>(RHSInputValues, 1) == kTfLiteOk);
160 CHECK(armnnInterpreter.Invoke() == kTfLiteOk);
161 std::vector<T> armnnOutputValues = armnnInterpreter.GetOutputResult<T>(0);
162 std::vector<int32_t> armnnOutputShape = armnnInterpreter.GetOutputShape(0);
163
164 armnnDelegate::CompareOutputData<T>(tfLiteOutputValues, armnnOutputValues, expectedOutputValues);
165 armnnDelegate::CompareOutputShape(tfLiteOutputShape, armnnOutputShape, outputShape);
166
167 tfLiteInterpreter.Cleanup();
168 armnnInterpreter.Cleanup();
169 }
170
171 } // anonymous namespace
172
173
174
175
176