XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #pragma once |
| 10 | |
| 11 | #include <gtest/gtest.h> |
| 12 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 13 | #include <algorithm> |
| 14 | #include <cfloat> |
| 15 | #include <cmath> |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 16 | #include <cstddef> |
| 17 | #include <cstdlib> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 18 | #include <functional> |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 19 | #include <limits> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 20 | #include <random> |
| 21 | #include <vector> |
| 22 | |
| 23 | #include <xnnpack/params.h> |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 24 | #include <xnnpack/requantization-stubs.h> |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 25 | #include <xnnpack/requantization.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 26 | |
| 27 | |
| 28 | class RequantizationTester { |
| 29 | public: |
| 30 | inline RequantizationTester& s(uint32_t s) { |
| 31 | this->s_ = s; |
| 32 | return *this; |
| 33 | } |
| 34 | |
| 35 | inline uint32_t s() const { |
| 36 | return this->s_; |
| 37 | } |
| 38 | |
| 39 | inline float scale() const { |
| 40 | return ldexpf(1.0f, -s()); |
| 41 | } |
| 42 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 43 | inline RequantizationTester& zero_point(int32_t zero_point) { |
| 44 | this->zero_point_ = zero_point; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 45 | return *this; |
| 46 | } |
| 47 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 48 | inline int32_t zero_point() const { |
| 49 | return this->zero_point_; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 50 | } |
| 51 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 52 | inline RequantizationTester& qmin(int16_t qmin) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 53 | this->qmin_ = qmin; |
| 54 | return *this; |
| 55 | } |
| 56 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 57 | inline int16_t qmin() const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 58 | return this->qmin_; |
| 59 | } |
| 60 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 61 | inline RequantizationTester& qmax(int16_t qmax) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 62 | this->qmax_ = qmax; |
| 63 | return *this; |
| 64 | } |
| 65 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 66 | inline int16_t qmax() const { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 67 | return this->qmax_; |
| 68 | } |
| 69 | |
| 70 | inline RequantizationTester& iterations(size_t iterations) { |
| 71 | this->iterations_ = iterations; |
| 72 | return *this; |
| 73 | } |
| 74 | |
| 75 | inline size_t iterations() const { |
| 76 | return this->iterations_; |
| 77 | } |
| 78 | |
| 79 | /* |
| 80 | * Test that requantization of numbers ((i - zero point) * 2**s) with |
| 81 | * - scale = exp2(-s) |
| 82 | * - zero point in [0, 255] |
| 83 | * - no output clamping |
| 84 | * produces exactly i, provided that ((i - zero point) * 2**s) does not overflow. |
| 85 | */ |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 86 | void TestExactDivideByPO2(xnn_qu8_requantization_function requantize) const { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 87 | ASSERT_GE(zero_point(), std::numeric_limits<uint8_t>::min()); |
| 88 | ASSERT_LE(zero_point(), std::numeric_limits<uint8_t>::max()); |
| 89 | ASSERT_GE(qmin(), std::numeric_limits<uint8_t>::min()); |
| 90 | ASSERT_LE(qmin(), std::numeric_limits<uint8_t>::max()); |
| 91 | ASSERT_GE(qmax(), std::numeric_limits<uint8_t>::min()); |
| 92 | ASSERT_LE(qmax(), std::numeric_limits<uint8_t>::max()); |
| 93 | ASSERT_LT(qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 94 | |
| 95 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 96 | ASSERT_GE(s(), 1); |
| 97 | ASSERT_LT(s(), 32); |
| 98 | |
| 99 | std::vector<int32_t> inputs(256); |
| 100 | std::vector<uint8_t> outputs(inputs.size()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 101 | const int32_t max_i = (uint32_t(std::numeric_limits<int32_t>::max()) >> s()) + zero_point(); |
| 102 | const int32_t min_i = -(-uint32_t(std::numeric_limits<int32_t>::min()) >> s()) + zero_point(); |
| 103 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
| 104 | const int32_t clamped_i = std::max(min_i, std::min(max_i, i)); |
| 105 | inputs[i] = int32_t(uint32_t(clamped_i - zero_point()) << s()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 106 | } |
| 107 | requantize(inputs.size(), inputs.data(), |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 108 | scale(), zero_point(), qmin(), qmax(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 109 | outputs.data()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 110 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
| 111 | const int32_t clamped_i = std::max(min_i, std::min(max_i, i)); |
| 112 | ASSERT_EQ(uint32_t(clamped_i), uint32_t(outputs[i])) |
Marat Dukhan | 138560c | 2020-08-03 18:57:34 -0700 | [diff] [blame] | 113 | << "i = " << i << ", clamped i = " << clamped_i << ", input = " << inputs[i] |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 114 | << ", min i = " << min_i << ", max i = " << max_i |
| 115 | << ", s = " << s() << ", zero point = " << zero_point(); |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | /* |
| 120 | * Test that requantization of numbers ((i - zero point) * 2**s) with |
| 121 | * - scale = exp2(-s) |
| 122 | * - zero point in [-128, 127] |
| 123 | * - no output clamping |
| 124 | * produces exactly i, provided that ((i - zero point) * 2**s) does not overflow. |
| 125 | */ |
| 126 | void TestExactDivideByPO2(xnn_qs8_requantization_function requantize) const { |
| 127 | ASSERT_GE(zero_point(), std::numeric_limits<int8_t>::min()); |
| 128 | ASSERT_LE(zero_point(), std::numeric_limits<int8_t>::max()); |
| 129 | ASSERT_GE(qmin(), std::numeric_limits<int8_t>::min()); |
| 130 | ASSERT_LE(qmin(), std::numeric_limits<int8_t>::max()); |
| 131 | ASSERT_GE(qmax(), std::numeric_limits<int8_t>::min()); |
| 132 | ASSERT_LE(qmax(), std::numeric_limits<int8_t>::max()); |
| 133 | ASSERT_LT(qmin(), qmax()); |
| 134 | |
| 135 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 136 | ASSERT_GE(s(), 1); |
| 137 | ASSERT_LT(s(), 32); |
| 138 | |
| 139 | std::vector<int32_t> inputs(256); |
| 140 | std::vector<int8_t> outputs(inputs.size()); |
| 141 | const int32_t max_i = (uint32_t(std::numeric_limits<int32_t>::max()) >> s()) + zero_point(); |
| 142 | const int32_t min_i = -(-uint32_t(std::numeric_limits<int32_t>::min()) >> s()) + zero_point(); |
| 143 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 144 | const int32_t clamped_i = std::max(min_i, std::min(max_i, i)); |
| 145 | inputs[i - std::numeric_limits<int8_t>::min()] = int32_t(uint32_t(clamped_i - zero_point()) << s()); |
| 146 | } |
| 147 | requantize(inputs.size(), inputs.data(), |
| 148 | scale(), zero_point(), qmin(), qmax(), |
| 149 | outputs.data()); |
| 150 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 151 | const int32_t clamped_i = std::max(min_i, std::min(max_i, i)); |
| 152 | ASSERT_EQ(clamped_i, int32_t(outputs[i - std::numeric_limits<int8_t>::min()])) |
| 153 | << "i = " << i << ", clamped i = " << clamped_i |
Marat Dukhan | 138560c | 2020-08-03 18:57:34 -0700 | [diff] [blame] | 154 | << ", input = " << inputs[i - std::numeric_limits<int8_t>::min()] |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 155 | << ", min i = " << min_i << ", max i = " << max_i |
| 156 | << ", s = " << s() << ", zero point = " << zero_point(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 157 | } |
| 158 | } |
| 159 | |
| 160 | /* |
| 161 | * Test that requantization of numbers (i * 2**s + sign(i - zero point) * 2**(s-1)) with |
| 162 | * - scale = exp2(-s) |
| 163 | * - zero point in [1, 255] |
| 164 | * - no output clamping |
| 165 | * produces exactly i, provided that ((i - zero point) * 2**s) does not overflow. |
| 166 | */ |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 167 | void TestDivideByPO2WithRoundingUp(xnn_qu8_requantization_function requantize) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 168 | ASSERT_GE(zero_point(), std::numeric_limits<uint8_t>::min()); |
| 169 | ASSERT_LE(zero_point(), std::numeric_limits<uint8_t>::max()); |
| 170 | ASSERT_GE(qmin(), std::numeric_limits<uint8_t>::min()); |
| 171 | ASSERT_LE(qmin(), std::numeric_limits<uint8_t>::max()); |
| 172 | ASSERT_GE(qmax(), std::numeric_limits<uint8_t>::min()); |
| 173 | ASSERT_LE(qmax(), std::numeric_limits<uint8_t>::max()); |
| 174 | ASSERT_LT(qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 175 | |
| 176 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 177 | ASSERT_GE(s(), 1); |
| 178 | ASSERT_LT(s(), 32); |
| 179 | |
| 180 | std::vector<int32_t> inputs(256); |
| 181 | std::vector<uint8_t> outputs(inputs.size()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 182 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 183 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) - |
| 184 | (INT64_C(1) << (s() - 1)) + (int64_t) (i <= zero_point()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 185 | inputs[i] = int32_t(input); |
| 186 | } |
| 187 | requantize(inputs.size(), inputs.data(), |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 188 | scale(), zero_point(), qmin(), qmax(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 189 | outputs.data()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 190 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 191 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) - |
| 192 | (INT64_C(1) << (s() - 1)) + (int64_t) (i <= zero_point()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 193 | if (int32_t(input) == input) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 194 | ASSERT_EQ(i, int32_t(outputs[i])) |
| 195 | << "i = " << i << ", input = " << input |
| 196 | << ", s = " << s() << ", zero point = " << zero_point(); |
| 197 | } |
| 198 | } |
| 199 | } |
| 200 | |
| 201 | /* |
| 202 | * Test that requantization of numbers (i * 2**s + sign(i - zero point) * 2**(s-1)) with |
| 203 | * - scale = exp2(-s) |
| 204 | * - zero point in [-128, 127] |
| 205 | * - no output clamping |
| 206 | * produces exactly i, provided that ((i - zero point) * 2**s) does not overflow. |
| 207 | */ |
| 208 | void TestDivideByPO2WithRoundingUp(xnn_qs8_requantization_function requantize) { |
| 209 | ASSERT_GE(zero_point(), std::numeric_limits<int8_t>::min()); |
| 210 | ASSERT_LE(zero_point(), std::numeric_limits<int8_t>::max()); |
| 211 | ASSERT_GE(qmin(), std::numeric_limits<int8_t>::min()); |
| 212 | ASSERT_LE(qmin(), std::numeric_limits<int8_t>::max()); |
| 213 | ASSERT_GE(qmax(), std::numeric_limits<int8_t>::min()); |
| 214 | ASSERT_LE(qmax(), std::numeric_limits<int8_t>::max()); |
| 215 | ASSERT_LT(qmin(), qmax()); |
| 216 | |
| 217 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 218 | ASSERT_GE(s(), 1); |
| 219 | ASSERT_LT(s(), 32); |
| 220 | |
| 221 | std::vector<int32_t> inputs(256); |
| 222 | std::vector<int8_t> outputs(inputs.size()); |
| 223 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 224 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) - |
| 225 | (INT64_C(1) << (s() - 1)) + (int64_t) (i <= zero_point()); |
| 226 | inputs[i - std::numeric_limits<int8_t>::min()] = int32_t(input); |
| 227 | } |
| 228 | requantize(inputs.size(), inputs.data(), |
| 229 | scale(), zero_point(), qmin(), qmax(), |
| 230 | outputs.data()); |
| 231 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 232 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) - |
| 233 | (INT64_C(1) << (s() - 1)) + (int64_t) (i <= zero_point()); |
| 234 | if (int32_t(input) == input) { |
| 235 | ASSERT_EQ(i, int32_t(outputs[i - std::numeric_limits<int8_t>::min()])) |
| 236 | << "i = " << i << ", input = " << input |
| 237 | << ", s = " << s() << ", zero point = " << zero_point(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 238 | } |
| 239 | } |
| 240 | } |
| 241 | |
| 242 | /* |
| 243 | * Test that requantization of numbers (i * 2**s + sign(i - zero point) * 2**(s-1)) with |
| 244 | * - scale = exp2(-s) |
| 245 | * - zero point in [1, 255] |
| 246 | * - no output clamping |
| 247 | * produces exactly i, provided that ((i - zero point) * 2**s) does not overflow. |
| 248 | */ |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 249 | void TestDivideByPO2WithRoundingDown(xnn_qu8_requantization_function requantize) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 250 | ASSERT_GE(zero_point(), std::numeric_limits<uint8_t>::min()); |
| 251 | ASSERT_LE(zero_point(), std::numeric_limits<uint8_t>::max()); |
| 252 | ASSERT_GE(qmin(), std::numeric_limits<uint8_t>::min()); |
| 253 | ASSERT_LE(qmin(), std::numeric_limits<uint8_t>::max()); |
| 254 | ASSERT_GE(qmax(), std::numeric_limits<uint8_t>::min()); |
| 255 | ASSERT_LE(qmax(), std::numeric_limits<uint8_t>::max()); |
| 256 | ASSERT_LT(qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 257 | |
| 258 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 259 | ASSERT_GE(s(), 1); |
| 260 | ASSERT_LT(s(), 32); |
| 261 | |
| 262 | std::vector<int32_t> inputs(256); |
| 263 | std::vector<uint8_t> outputs(inputs.size()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 264 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 265 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) + |
| 266 | (INT64_C(1) << (s() - 1)) - (int64_t) (i >= zero_point()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 267 | inputs[i] = int32_t(input); |
| 268 | } |
| 269 | requantize(inputs.size(), inputs.data(), |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 270 | scale(), zero_point(), qmin(), qmax(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 271 | outputs.data()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 272 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 273 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) + |
| 274 | (INT64_C(1) << (s() - 1)) - (int64_t) (i >= zero_point()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 275 | if (int32_t(input) == input) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 276 | ASSERT_EQ(i, int32_t(outputs[i])) |
| 277 | << "i = " << i << ", input = " << input |
| 278 | << ", s = " << s() << ", zero point = " << zero_point(); |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | |
| 283 | /* |
| 284 | * Test that requantization of numbers (i * 2**s + sign(i - zero point) * 2**(s-1)) with |
| 285 | * - scale = exp2(-s) |
| 286 | * - zero point in [-128, 127] |
| 287 | * - no output clamping |
| 288 | * produces exactly i, provided that ((i - zero point) * 2**s) does not overflow. |
| 289 | */ |
| 290 | void TestDivideByPO2WithRoundingDown(xnn_qs8_requantization_function requantize) { |
| 291 | ASSERT_GE(zero_point(), std::numeric_limits<int8_t>::min()); |
| 292 | ASSERT_LE(zero_point(), std::numeric_limits<int8_t>::max()); |
| 293 | ASSERT_GE(qmin(), std::numeric_limits<int8_t>::min()); |
| 294 | ASSERT_LE(qmin(), std::numeric_limits<int8_t>::max()); |
| 295 | ASSERT_GE(qmax(), std::numeric_limits<int8_t>::min()); |
| 296 | ASSERT_LE(qmax(), std::numeric_limits<int8_t>::max()); |
| 297 | ASSERT_LT(qmin(), qmax()); |
| 298 | |
| 299 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 300 | ASSERT_GE(s(), 1); |
| 301 | ASSERT_LT(s(), 32); |
| 302 | |
| 303 | std::vector<int32_t> inputs(256); |
| 304 | std::vector<int8_t> outputs(inputs.size()); |
| 305 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 306 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) + |
| 307 | (INT64_C(1) << (s() - 1)) - (int64_t) (i >= zero_point()); |
| 308 | inputs[i - std::numeric_limits<int8_t>::min()] = int32_t(input); |
| 309 | } |
| 310 | requantize(inputs.size(), inputs.data(), |
| 311 | scale(), zero_point(), qmin(), qmax(), |
| 312 | outputs.data()); |
| 313 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 314 | const int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()) + |
| 315 | (INT64_C(1) << (s() - 1)) - (int64_t) (i >= zero_point()); |
| 316 | if (int32_t(input) == input) { |
| 317 | ASSERT_EQ(i, int32_t(outputs[i - std::numeric_limits<int8_t>::min()])) |
| 318 | << "i = " << i << ", input = " << input |
| 319 | << ", s = " << s() << ", zero point = " << zero_point(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 320 | } |
| 321 | } |
| 322 | } |
| 323 | |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 324 | void TestDivideByPO2WithRoundingAway(xnn_qu8_requantization_function requantize) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 325 | ASSERT_GE(zero_point(), std::numeric_limits<uint8_t>::min()); |
| 326 | ASSERT_LE(zero_point(), std::numeric_limits<uint8_t>::max()); |
| 327 | ASSERT_GE(qmin(), std::numeric_limits<uint8_t>::min()); |
| 328 | ASSERT_LE(qmin(), std::numeric_limits<uint8_t>::max()); |
| 329 | ASSERT_GE(qmax(), std::numeric_limits<uint8_t>::min()); |
| 330 | ASSERT_LE(qmax(), std::numeric_limits<uint8_t>::max()); |
| 331 | ASSERT_LT(qmin(), qmax()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 332 | |
| 333 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 334 | ASSERT_GE(s(), 1); |
| 335 | ASSERT_LT(s(), 32); |
| 336 | |
| 337 | std::vector<int32_t> inputs(256); |
| 338 | std::vector<uint8_t> outputs(inputs.size()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 339 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 340 | int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 341 | if (input > 0) { |
| 342 | input -= INT64_C(1) << (s() - 1); |
| 343 | } else if (input < 0) { |
| 344 | input += INT64_C(1) << (s() - 1); |
| 345 | } |
| 346 | inputs[i] = int32_t(input); |
| 347 | } |
| 348 | requantize(inputs.size(), inputs.data(), |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 349 | scale(), zero_point(), qmin(), qmax(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 350 | outputs.data()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 351 | for (int32_t i = 0; i <= std::numeric_limits<uint8_t>::max(); i++) { |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 352 | int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 353 | if (input > 0) { |
| 354 | input -= INT64_C(1) << (s() - 1); |
| 355 | } else if (input < 0) { |
| 356 | input += INT64_C(1) << (s() - 1); |
| 357 | } |
| 358 | if (int32_t(input) == input) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 359 | ASSERT_EQ(i, int32_t(outputs[i])) |
| 360 | << "i = " << i << ", input = " << input |
| 361 | << ", s = " << s() << ", zero point = " << zero_point(); |
| 362 | } |
| 363 | } |
| 364 | } |
| 365 | |
| 366 | void TestDivideByPO2WithRoundingAway(xnn_qs8_requantization_function requantize) { |
| 367 | ASSERT_GE(zero_point(), std::numeric_limits<int8_t>::min()); |
| 368 | ASSERT_LE(zero_point(), std::numeric_limits<int8_t>::max()); |
| 369 | ASSERT_GE(qmin(), std::numeric_limits<int8_t>::min()); |
| 370 | ASSERT_LE(qmin(), std::numeric_limits<int8_t>::max()); |
| 371 | ASSERT_GE(qmax(), std::numeric_limits<int8_t>::min()); |
| 372 | ASSERT_LE(qmax(), std::numeric_limits<int8_t>::max()); |
| 373 | ASSERT_LT(qmin(), qmax()); |
| 374 | |
| 375 | /* Note: need s >= 1 to ensure scale = exp2(-s) < 1.0 */ |
| 376 | ASSERT_GE(s(), 1); |
| 377 | ASSERT_LT(s(), 32); |
| 378 | |
| 379 | std::vector<int32_t> inputs(256); |
| 380 | std::vector<int8_t> outputs(inputs.size()); |
| 381 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 382 | int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()); |
| 383 | if (input > 0) { |
| 384 | input -= INT64_C(1) << (s() - 1); |
| 385 | } else if (input < 0) { |
| 386 | input += INT64_C(1) << (s() - 1); |
| 387 | } |
| 388 | inputs[i - std::numeric_limits<int8_t>::min()] = int32_t(input); |
| 389 | } |
| 390 | requantize(inputs.size(), inputs.data(), |
| 391 | scale(), zero_point(), qmin(), qmax(), |
| 392 | outputs.data()); |
| 393 | for (int32_t i = std::numeric_limits<int8_t>::min(); i <= std::numeric_limits<int8_t>::max(); i++) { |
| 394 | int64_t input = RequantizationTester::ShiftLeft(i - zero_point(), s()); |
| 395 | if (input > 0) { |
| 396 | input -= INT64_C(1) << (s() - 1); |
| 397 | } else if (input < 0) { |
| 398 | input += INT64_C(1) << (s() - 1); |
| 399 | } |
| 400 | if (int32_t(input) == input) { |
| 401 | ASSERT_EQ(i, int32_t(outputs[i - std::numeric_limits<int8_t>::min()])) |
| 402 | << "i = " << i << ", input = " << input |
| 403 | << ", s = " << s() << ", zero point = " << zero_point(); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 404 | } |
| 405 | } |
| 406 | } |
| 407 | |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 408 | void TestSpecialCases(xnn_qu8_requantization_function requantize) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 409 | ASSERT_GE(qmin(), std::numeric_limits<uint8_t>::min()); |
| 410 | ASSERT_LE(qmin(), std::numeric_limits<uint8_t>::max()); |
| 411 | ASSERT_GE(qmax(), std::numeric_limits<uint8_t>::min()); |
| 412 | ASSERT_LE(qmax(), std::numeric_limits<uint8_t>::max()); |
| 413 | ASSERT_LT(qmin(), qmax()); |
| 414 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 415 | std::vector<int32_t> inputs(256); |
| 416 | std::vector<uint8_t> outputs(inputs.size()); |
| 417 | |
| 418 | std::fill(inputs.begin(), inputs.end(), std::numeric_limits<int32_t>::min()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 419 | for (int32_t zero_point = 0; zero_point <= std::numeric_limits<uint8_t>::max(); zero_point++) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 420 | requantize( |
| 421 | inputs.size(), |
| 422 | inputs.data(), |
| 423 | ldexpf(1.0f, -32) /* scale */, |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 424 | zero_point /* zero point */, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 425 | std::numeric_limits<uint8_t>::min(), |
| 426 | std::numeric_limits<uint8_t>::max(), |
| 427 | outputs.data()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 428 | for (size_t i = 0; i < outputs.size(); i++) { |
| 429 | ASSERT_EQ(std::max(int32_t(int32_t(std::numeric_limits<uint8_t>::min())), zero_point - 1), int32_t(outputs[i])); |
| 430 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 431 | } |
| 432 | |
| 433 | std::fill(inputs.begin(), inputs.end(), std::numeric_limits<int32_t>::max()); |
| 434 | requantize( |
| 435 | inputs.size(), |
| 436 | inputs.data(), |
| 437 | 0x1.FFFFFEp-1f /* scale */, |
| 438 | std::numeric_limits<uint8_t>::max() /* zero point */, |
| 439 | std::numeric_limits<uint8_t>::min(), |
| 440 | std::numeric_limits<uint8_t>::max(), |
| 441 | outputs.data()); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 442 | for (size_t i = 0; i < outputs.size(); i++) { |
| 443 | ASSERT_EQ(std::numeric_limits<uint8_t>::max(), int32_t(outputs[i])); |
| 444 | } |
| 445 | } |
| 446 | |
| 447 | void TestSpecialCases(xnn_qs8_requantization_function requantize) { |
| 448 | ASSERT_GE(qmin(), std::numeric_limits<int8_t>::min()); |
| 449 | ASSERT_LE(qmin(), std::numeric_limits<int8_t>::max()); |
| 450 | ASSERT_GE(qmax(), std::numeric_limits<int8_t>::min()); |
| 451 | ASSERT_LE(qmax(), std::numeric_limits<int8_t>::max()); |
| 452 | ASSERT_LT(qmin(), qmax()); |
| 453 | |
| 454 | std::vector<int32_t> inputs(256); |
| 455 | std::vector<int8_t> outputs(inputs.size()); |
| 456 | |
| 457 | std::fill(inputs.begin(), inputs.end(), std::numeric_limits<int32_t>::min()); |
| 458 | for (int32_t zero_point = std::numeric_limits<int8_t>::min(); |
| 459 | zero_point <= std::numeric_limits<int8_t>::max(); |
| 460 | zero_point++) |
| 461 | { |
| 462 | requantize( |
| 463 | inputs.size(), |
| 464 | inputs.data(), |
| 465 | ldexpf(1.0f, -32) /* scale */, |
| 466 | zero_point, |
| 467 | std::numeric_limits<int8_t>::min(), |
| 468 | std::numeric_limits<int8_t>::max(), |
| 469 | outputs.data()); |
| 470 | for (size_t i = 0; i < outputs.size(); i++) { |
| 471 | ASSERT_EQ(std::max(int32_t(std::numeric_limits<int8_t>::min()), zero_point - 1), int32_t(outputs[i])); |
| 472 | } |
| 473 | } |
| 474 | |
| 475 | std::fill(inputs.begin(), inputs.end(), std::numeric_limits<int32_t>::max()); |
| 476 | requantize( |
| 477 | inputs.size(), |
| 478 | inputs.data(), |
| 479 | 0x1.FFFFFEp-1f /* scale */, |
| 480 | std::numeric_limits<int8_t>::max() /* zero point */, |
| 481 | std::numeric_limits<int8_t>::min(), |
| 482 | std::numeric_limits<int8_t>::max(), |
| 483 | outputs.data()); |
| 484 | for (size_t i = 0; i < outputs.size(); i++) { |
| 485 | ASSERT_EQ(std::numeric_limits<int8_t>::max(), int32_t(outputs[i])); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 486 | } |
| 487 | } |
| 488 | |
Marat Dukhan | 5b69f8b | 2020-07-24 15:26:48 -0700 | [diff] [blame] | 489 | void TestRandomCasesPrecise(xnn_qu8_requantization_function requantize) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 490 | ASSERT_GE(zero_point(), std::numeric_limits<uint8_t>::min()); |
| 491 | ASSERT_LE(zero_point(), std::numeric_limits<uint8_t>::max()); |
| 492 | ASSERT_GE(qmin(), std::numeric_limits<uint8_t>::min()); |
| 493 | ASSERT_LE(qmin(), std::numeric_limits<uint8_t>::max()); |
| 494 | ASSERT_GE(qmax(), std::numeric_limits<uint8_t>::min()); |
| 495 | ASSERT_LE(qmax(), std::numeric_limits<uint8_t>::max()); |
| 496 | ASSERT_LT(qmin(), qmax()); |
| 497 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 498 | std::random_device random_device; |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 499 | std::mt19937 rng(random_device()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 500 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 501 | auto u8rng = |
| 502 | std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 503 | |
| 504 | std::vector<int32_t> inputs(4096); |
| 505 | std::vector<uint8_t> outputs(inputs.size()); |
| 506 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 507 | std::uniform_real_distribution<float> scale_distribution(0x1.000000p-23f, 0x1.FFFFFEp-1f); |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 508 | const float scale = scale_distribution(rng); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 509 | for (size_t i = 0; i < inputs.size(); i++) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 510 | const uint8_t approximate_output = std::min(std::max(uint8_t(u8rng()), uint8_t(qmin())), uint8_t(qmax())); |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 511 | const int32_t input = int32_t(double(approximate_output) / double(scale)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 512 | inputs[i] = input; |
| 513 | } |
| 514 | |
| 515 | requantize( |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 516 | inputs.size(), inputs.data(), scale, zero_point(), qmin(), qmax(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 517 | outputs.data()); |
| 518 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 519 | /* Ensure that outputs are not all identical, as in this case the test doesn't validate much */ |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 520 | ASSERT_NE( |
| 521 | *std::max_element(outputs.cbegin(), outputs.cend()), |
| 522 | *std::min_element(outputs.cbegin(), outputs.cend())); |
| 523 | |
| 524 | for (size_t i = 0; i < inputs.size(); i++) { |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 525 | const uint8_t reference_output = |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 526 | xnn_qu8_requantize_precise(inputs[i], scale, zero_point(), qmin(), qmax()); |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 527 | ASSERT_EQ(uint32_t(reference_output), uint32_t(outputs[i])); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 528 | } |
| 529 | } |
| 530 | } |
| 531 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 532 | void TestRandomCasesPrecise(xnn_qs8_requantization_function requantize) { |
| 533 | ASSERT_GE(zero_point(), std::numeric_limits<int8_t>::min()); |
| 534 | ASSERT_LE(zero_point(), std::numeric_limits<int8_t>::max()); |
| 535 | ASSERT_GE(qmin(), std::numeric_limits<int8_t>::min()); |
| 536 | ASSERT_LE(qmin(), std::numeric_limits<int8_t>::max()); |
| 537 | ASSERT_GE(qmax(), std::numeric_limits<int8_t>::min()); |
| 538 | ASSERT_LE(qmax(), std::numeric_limits<int8_t>::max()); |
| 539 | ASSERT_LT(qmin(), qmax()); |
| 540 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 541 | std::random_device random_device; |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 542 | std::mt19937 rng(random_device()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 543 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 544 | auto i8rng = std::bind( |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 545 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), std::ref(rng)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 546 | |
| 547 | std::vector<int32_t> inputs(4096); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 548 | std::vector<int8_t> outputs(inputs.size()); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 549 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 550 | std::uniform_real_distribution<float> scale_distribution(0x1.000000p-23f, 0x1.FFFFFEp-1f); |
Marat Dukhan | 5ce30d9 | 2020-04-14 03:31:26 -0700 | [diff] [blame] | 551 | const float scale = scale_distribution(rng); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 552 | for (size_t i = 0; i < inputs.size(); i++) { |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 553 | const int8_t approximate_output = std::min(std::max(int8_t(i8rng()), int8_t(qmin())), int8_t(qmax())); |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 554 | const int32_t input = int32_t(double(approximate_output) / double(scale)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 555 | inputs[i] = input; |
| 556 | } |
| 557 | |
| 558 | requantize( |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 559 | inputs.size(), inputs.data(), scale, zero_point(), qmin(), qmax(), |
| 560 | outputs.data()); |
| 561 | |
| 562 | /* Ensure that outputs are not all identical, as in this case the test doesn't validate much */ |
| 563 | ASSERT_NE( |
| 564 | *std::max_element(outputs.cbegin(), outputs.cend()), |
| 565 | *std::min_element(outputs.cbegin(), outputs.cend())); |
| 566 | |
| 567 | for (size_t i = 0; i < inputs.size(); i++) { |
| 568 | const int8_t reference_output = |
| 569 | xnn_qs8_requantize_precise(inputs[i], scale, zero_point(), qmin(), qmax()); |
| 570 | ASSERT_EQ(int32_t(reference_output), int32_t(outputs[i])); |
| 571 | } |
| 572 | } |
| 573 | } |
| 574 | |
| 575 | void TestRandomCasesApproximate(xnn_qu8_requantization_function requantize) { |
| 576 | ASSERT_GE(zero_point(), std::numeric_limits<uint8_t>::min()); |
| 577 | ASSERT_LE(zero_point(), std::numeric_limits<uint8_t>::max()); |
| 578 | ASSERT_GE(qmin(), std::numeric_limits<uint8_t>::min()); |
| 579 | ASSERT_LE(qmin(), std::numeric_limits<uint8_t>::max()); |
| 580 | ASSERT_GE(qmax(), std::numeric_limits<uint8_t>::min()); |
| 581 | ASSERT_LE(qmax(), std::numeric_limits<uint8_t>::max()); |
| 582 | ASSERT_LT(qmin(), qmax()); |
| 583 | |
| 584 | std::random_device random_device; |
| 585 | std::mt19937 rng(random_device()); |
| 586 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 587 | auto u8rng = |
| 588 | std::bind(std::uniform_int_distribution<uint32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
| 589 | |
| 590 | std::vector<int32_t> inputs(4096); |
| 591 | std::vector<uint8_t> outputs(inputs.size()); |
| 592 | |
| 593 | std::uniform_real_distribution<float> scale_distribution(0x1.000000p-23f, 0x1.FFFFFEp-1f); |
| 594 | const float scale = scale_distribution(rng); |
| 595 | for (size_t i = 0; i < inputs.size(); i++) { |
| 596 | const uint8_t approximate_output = std::min(std::max(uint8_t(u8rng()), uint8_t(qmin())), uint8_t(qmax())); |
| 597 | const int32_t input = int32_t(double(approximate_output) / double(scale)); |
| 598 | inputs[i] = input; |
| 599 | } |
| 600 | |
| 601 | requantize( |
| 602 | inputs.size(), inputs.data(), scale, zero_point(), qmin(), qmax(), |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 603 | outputs.data()); |
| 604 | |
| 605 | /* Ensure that outputs are not all identical, as in this case Test doesn't validate much */ |
| 606 | ASSERT_NE( |
| 607 | *std::max_element(outputs.cbegin(), outputs.cend()), |
| 608 | *std::min_element(outputs.cbegin(), outputs.cend())); |
| 609 | |
| 610 | for (size_t i = 0; i < inputs.size(); i++) { |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 611 | const double reference_output = RequantizationTester::RequantizeApproximate( |
| 612 | inputs[i], scale, uint8_t(zero_point()), uint8_t(qmin()), uint8_t(qmax())); |
| 613 | ASSERT_LE(std::abs(reference_output - double(outputs[i])), 0.55) |
| 614 | << "input = " << inputs[i] << ", output = " << int32_t(outputs[i]) |
| 615 | << ", reference output = " << reference_output; |
| 616 | } |
| 617 | } |
| 618 | } |
| 619 | |
| 620 | void TestRandomCasesApproximate(xnn_qs8_requantization_function requantize) { |
| 621 | ASSERT_GE(zero_point(), std::numeric_limits<int8_t>::min()); |
| 622 | ASSERT_LE(zero_point(), std::numeric_limits<int8_t>::max()); |
| 623 | ASSERT_GE(qmin(), std::numeric_limits<int8_t>::min()); |
| 624 | ASSERT_LE(qmin(), std::numeric_limits<int8_t>::max()); |
| 625 | ASSERT_GE(qmax(), std::numeric_limits<int8_t>::min()); |
| 626 | ASSERT_LE(qmax(), std::numeric_limits<int8_t>::max()); |
| 627 | ASSERT_LT(qmin(), qmax()); |
| 628 | |
| 629 | std::random_device random_device; |
| 630 | std::mt19937 rng(random_device()); |
| 631 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 632 | auto i8rng = std::bind( |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 633 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), std::ref(rng)); |
| 634 | |
| 635 | std::vector<int32_t> inputs(4096); |
| 636 | std::vector<int8_t> outputs(inputs.size()); |
| 637 | |
| 638 | std::uniform_real_distribution<float> scale_distribution(0x1.000000p-23f, 0x1.FFFFFEp-1f); |
| 639 | const float scale = scale_distribution(rng); |
| 640 | for (size_t i = 0; i < inputs.size(); i++) { |
Marat Dukhan | ecd8311 | 2020-08-03 21:50:28 -0700 | [diff] [blame] | 641 | const int8_t approximate_output = std::min(std::max(int8_t(i8rng()), int8_t(qmin())), int8_t(qmax())); |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 642 | const int32_t input = int32_t(double(approximate_output) / double(scale)); |
| 643 | inputs[i] = input; |
| 644 | } |
| 645 | |
| 646 | requantize( |
| 647 | inputs.size(), inputs.data(), scale, zero_point(), qmin(), qmax(), |
| 648 | outputs.data()); |
| 649 | |
| 650 | /* Ensure that outputs are not all identical, as in this case Test doesn't validate much */ |
| 651 | ASSERT_NE( |
| 652 | *std::max_element(outputs.cbegin(), outputs.cend()), |
| 653 | *std::min_element(outputs.cbegin(), outputs.cend())); |
| 654 | |
| 655 | for (size_t i = 0; i < inputs.size(); i++) { |
| 656 | const double reference_output = RequantizationTester::RequantizeApproximate( |
| 657 | inputs[i], scale, int8_t(zero_point()), int8_t(qmin()), int8_t(qmax())); |
| 658 | ASSERT_LE(std::abs(reference_output - double(outputs[i])), 0.55) |
| 659 | << "input = " << inputs[i] << ", output = " << int32_t(outputs[i]) |
| 660 | << ", reference output = " << reference_output; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 661 | } |
| 662 | } |
| 663 | } |
| 664 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 665 | static inline int64_t ShiftLeft(int64_t w, uint32_t n) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 666 | return (int64_t) ((uint64_t) w << n); |
| 667 | } |
| 668 | |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 669 | static inline double RequantizeApproximate( |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 670 | int32_t value, |
| 671 | float scale, |
Marat Dukhan | fe7acb6 | 2020-03-09 19:30:05 -0700 | [diff] [blame] | 672 | uint8_t zero_point, |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 673 | uint8_t qmin, |
| 674 | uint8_t qmax) |
| 675 | { |
| 676 | assert(scale < 1.0f); |
| 677 | assert(scale >= 0x1.0p-32f); |
| 678 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 679 | return std::min(std::max(double(value) * double(scale) + double(zero_point), double(qmin)), double(qmax)); |
| 680 | } |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 681 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 682 | static inline double RequantizeApproximate( |
| 683 | int32_t value, |
| 684 | float scale, |
| 685 | int8_t zero_point, |
| 686 | int8_t qmin, |
| 687 | int8_t qmax) |
| 688 | { |
| 689 | assert(scale < 1.0f); |
| 690 | assert(scale >= 0x1.0p-32f); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 691 | |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 692 | return std::min(std::max(double(value) * double(scale) + double(zero_point), double(qmin)), double(qmax)); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 693 | } |
| 694 | |
| 695 | private: |
Marat Dukhan | 2e23d2b | 2020-07-29 16:01:37 -0700 | [diff] [blame] | 696 | uint32_t s_{1}; |
| 697 | int32_t zero_point_{0}; |
| 698 | int16_t qmin_{std::numeric_limits<int16_t>::min()}; |
| 699 | int16_t qmax_{std::numeric_limits<int16_t>::max()}; |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 700 | size_t iterations_{1}; |
| 701 | }; |