1 //
2 // Copyright © 2017 Arm Ltd and Contributors. All rights reserved.
3 // SPDX-License-Identifier: MIT
4 //
5 #pragma once
6
7 #include <armnnTestUtils/PredicateResult.hpp>
8
9 #include <armnn/Tensor.hpp>
10 #include <armnn/utility/Assert.hpp>
11 #include <armnnUtils/FloatingPointComparison.hpp>
12
13 #include <armnnUtils/QuantizeHelper.hpp>
14
15 #include <doctest/doctest.h>
16
17 #include <array>
18 #include <cmath>
19 #include <random>
20 #include <vector>
21
22 constexpr float g_FloatCloseToZeroTolerance = 1.0e-6f;
23
24 template<typename T, bool isQuantized = true>
25 struct SelectiveComparer
26 {
CompareSelectiveComparer27 static bool Compare(T a, T b)
28 {
29 return (std::max(a, b) - std::min(a, b)) <= 1;
30 }
31
32 };
33
34 template<typename T>
35 struct SelectiveComparer<T, false>
36 {
CompareSelectiveComparer37 static bool Compare(T a, T b)
38 {
39 // If a or b is zero, percent_tolerance does an exact match, so compare to a small, constant tolerance instead.
40 if (a == 0.0f || b == 0.0f)
41 {
42 return std::abs(a - b) <= g_FloatCloseToZeroTolerance;
43 }
44
45 if (std::isinf(a) && a == b)
46 {
47 return true;
48 }
49
50 if (std::isnan(a) && std::isnan(b))
51 {
52 return true;
53 }
54
55 // For unquantized floats we use a tolerance of 1%.
56 return armnnUtils::within_percentage_tolerance(a, b);
57 }
58 };
59
60 template<typename T>
SelectiveCompare(T a,T b)61 bool SelectiveCompare(T a, T b)
62 {
63 return SelectiveComparer<T, armnn::IsQuantizedType<T>()>::Compare(a, b);
64 };
65
66 template<typename T>
SelectiveCompareBoolean(T a,T b)67 bool SelectiveCompareBoolean(T a, T b)
68 {
69 return (((a == 0) && (b == 0)) || ((a != 0) && (b != 0)));
70 };
71
72 template <typename T>
CompareTensors(const std::vector<T> & actualData,const std::vector<T> & expectedData,const armnn::TensorShape & actualShape,const armnn::TensorShape & expectedShape,bool compareBoolean=false,bool isDynamic=false)73 armnn::PredicateResult CompareTensors(const std::vector<T>& actualData,
74 const std::vector<T>& expectedData,
75 const armnn::TensorShape& actualShape,
76 const armnn::TensorShape& expectedShape,
77 bool compareBoolean = false,
78 bool isDynamic = false)
79 {
80 if (actualData.size() != expectedData.size())
81 {
82 armnn::PredicateResult res(false);
83 res.Message() << "Different data size ["
84 << actualData.size()
85 << "!="
86 << expectedData.size()
87 << "]";
88 return res;
89 }
90
91 if (actualShape.GetNumDimensions() != expectedShape.GetNumDimensions())
92 {
93 armnn::PredicateResult res(false);
94 res.Message() << "Different number of dimensions ["
95 << actualShape.GetNumDimensions()
96 << "!="
97 << expectedShape.GetNumDimensions()
98 << "]";
99 return res;
100 }
101
102 if (actualShape.GetNumElements() != expectedShape.GetNumElements())
103 {
104 armnn::PredicateResult res(false);
105 res.Message() << "Different number of elements ["
106 << actualShape.GetNumElements()
107 << "!="
108 << expectedShape.GetNumElements()
109 << "]";
110 return res;
111 }
112
113 unsigned int numberOfDimensions = actualShape.GetNumDimensions();
114
115 if (!isDynamic)
116 {
117 // Checks they are same shape.
118 for (unsigned int i = 0; i < numberOfDimensions; ++i)
119 {
120 if (actualShape[i] != expectedShape[i])
121 {
122 armnn::PredicateResult res(false);
123 res.Message() << "Different shapes ["
124 << actualShape[i]
125 << "!="
126 << expectedShape[i]
127 << "]";
128 return res;
129 }
130 }
131 }
132
133 // Fun iteration over n dimensions.
134 std::vector<unsigned int> indices;
135 for (unsigned int i = 0; i < numberOfDimensions; i++)
136 {
137 indices.emplace_back(0);
138 }
139
140 std::stringstream errorString;
141 int numFailedElements = 0;
142 constexpr int maxReportedDifferences = 3;
143 unsigned int index = 0;
144
145 // Compare data element by element.
146 while (true)
147 {
148 bool comparison;
149 // As true for uint8_t is non-zero (1-255) we must have a dedicated compare for Booleans.
150 if(compareBoolean)
151 {
152 comparison = SelectiveCompareBoolean(actualData[index], expectedData[index]);
153 }
154 else
155 {
156 comparison = SelectiveCompare(actualData[index], expectedData[index]);
157 }
158
159 if (!comparison)
160 {
161 ++numFailedElements;
162
163 if (numFailedElements <= maxReportedDifferences)
164 {
165 if (numFailedElements >= 2)
166 {
167 errorString << ", ";
168 }
169 errorString << "[";
170 for (unsigned int i = 0; i < numberOfDimensions; ++i)
171 {
172 errorString << indices[i];
173 if (i != numberOfDimensions - 1)
174 {
175 errorString << ",";
176 }
177 }
178 errorString << "]";
179
180 errorString << " (" << +actualData[index] << " != " << +expectedData[index] << ")";
181 }
182 }
183
184 ++indices[numberOfDimensions - 1];
185 for (unsigned int i=numberOfDimensions-1; i>0; i--)
186 {
187 if (indices[i] == actualShape[i])
188 {
189 indices[i] = 0;
190 ++indices[i - 1];
191 }
192 }
193 if (indices[0] == actualShape[0])
194 {
195 break;
196 }
197
198 index++;
199 }
200
201 armnn::PredicateResult comparisonResult(true);
202 if (numFailedElements > 0)
203 {
204 comparisonResult.SetResult(false);
205 comparisonResult.Message() << numFailedElements << " different values at: ";
206 if (numFailedElements > maxReportedDifferences)
207 {
208 errorString << ", ... (and " << (numFailedElements - maxReportedDifferences) << " other differences)";
209 }
210 comparisonResult.Message() << errorString.str();
211 }
212
213 return comparisonResult;
214 }
215
216 template <typename T>
MakeRandomTensor(const armnn::TensorInfo & tensorInfo,unsigned int seed,float min=-10.0f,float max=10.0f)217 std::vector<T> MakeRandomTensor(const armnn::TensorInfo& tensorInfo,
218 unsigned int seed,
219 float min = -10.0f,
220 float max = 10.0f)
221 {
222 std::mt19937 gen(seed);
223 std::uniform_real_distribution<float> dist(min, max);
224
225 std::vector<float> init(tensorInfo.GetNumElements());
226 for (unsigned int i = 0; i < init.size(); i++)
227 {
228 init[i] = dist(gen);
229 }
230
231 const float qScale = tensorInfo.GetQuantizationScale();
232 const int32_t qOffset = tensorInfo.GetQuantizationOffset();
233
234 return armnnUtils::QuantizedVector<T>(init, qScale, qOffset);
235 }
236