Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 1 | // Copyright 2019 Google LLC |
| 2 | // |
| 3 | // This source code is licensed under the BSD-style license found in the |
| 4 | // LICENSE file in the root directory of this source tree. |
| 5 | |
| 6 | #pragma once |
| 7 | |
| 8 | #include <gtest/gtest.h> |
| 9 | |
| 10 | #include <algorithm> |
| 11 | #include <cassert> |
| 12 | #include <cstddef> |
| 13 | #include <cstdlib> |
| 14 | #include <functional> |
| 15 | #include <random> |
| 16 | #include <vector> |
| 17 | |
Marat Dukhan | fcfdd2d | 2021-06-29 18:57:02 -0700 | [diff] [blame] | 18 | #include <fp16.h> |
| 19 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 20 | #include <xnnpack.h> |
| 21 | #include <xnnpack/params-init.h> |
| 22 | #include <xnnpack/params.h> |
| 23 | |
| 24 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 25 | class VUnaryMicrokernelTester { |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 26 | public: |
| 27 | enum class OpType { |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 28 | ReLU, |
Marat Dukhan | eecf8fd | 2020-06-09 08:59:37 -0700 | [diff] [blame] | 29 | RoundToNearestEven, |
| 30 | RoundTowardsZero, |
| 31 | RoundUp, |
| 32 | RoundDown, |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 33 | }; |
| 34 | |
| 35 | enum class Variant { |
| 36 | Native, |
| 37 | Scalar, |
| 38 | }; |
| 39 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 40 | inline VUnaryMicrokernelTester& batch_size(size_t batch_size) { |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 41 | assert(batch_size != 0); |
| 42 | this->batch_size_ = batch_size; |
| 43 | return *this; |
| 44 | } |
| 45 | |
| 46 | inline size_t batch_size() const { |
| 47 | return this->batch_size_; |
| 48 | } |
| 49 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 50 | inline VUnaryMicrokernelTester& inplace(bool inplace) { |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 51 | this->inplace_ = inplace; |
| 52 | return *this; |
| 53 | } |
| 54 | |
| 55 | inline bool inplace() const { |
| 56 | return this->inplace_; |
| 57 | } |
| 58 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 59 | inline VUnaryMicrokernelTester& slope(float slope) { |
Marat Dukhan | 8cc7efe | 2020-06-10 16:24:27 -0700 | [diff] [blame] | 60 | this->slope_ = slope; |
| 61 | return *this; |
| 62 | } |
| 63 | |
| 64 | inline float slope() const { |
| 65 | return this->slope_; |
| 66 | } |
| 67 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 68 | inline VUnaryMicrokernelTester& prescale(float prescale) { |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 69 | this->prescale_ = prescale; |
| 70 | return *this; |
| 71 | } |
| 72 | |
| 73 | inline float prescale() const { |
| 74 | return this->prescale_; |
| 75 | } |
| 76 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 77 | inline VUnaryMicrokernelTester& alpha(float alpha) { |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 78 | this->alpha_ = alpha; |
| 79 | return *this; |
| 80 | } |
| 81 | |
| 82 | inline float alpha() const { |
| 83 | return this->alpha_; |
| 84 | } |
| 85 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 86 | inline VUnaryMicrokernelTester& beta(float beta) { |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 87 | this->beta_ = beta; |
| 88 | return *this; |
| 89 | } |
| 90 | |
| 91 | inline float beta() const { |
| 92 | return this->beta_; |
| 93 | } |
| 94 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 95 | inline VUnaryMicrokernelTester& qmin(uint8_t qmin) { |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 96 | this->qmin_ = qmin; |
| 97 | return *this; |
| 98 | } |
| 99 | |
| 100 | inline uint8_t qmin() const { |
| 101 | return this->qmin_; |
| 102 | } |
| 103 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 104 | inline VUnaryMicrokernelTester& qmax(uint8_t qmax) { |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 105 | this->qmax_ = qmax; |
| 106 | return *this; |
| 107 | } |
| 108 | |
| 109 | inline uint8_t qmax() const { |
| 110 | return this->qmax_; |
| 111 | } |
| 112 | |
Marat Dukhan | 87ed45c | 2021-05-13 12:25:22 -0700 | [diff] [blame] | 113 | inline VUnaryMicrokernelTester& iterations(size_t iterations) { |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 114 | this->iterations_ = iterations; |
| 115 | return *this; |
| 116 | } |
| 117 | |
| 118 | inline size_t iterations() const { |
| 119 | return this->iterations_; |
| 120 | } |
| 121 | |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 122 | void Test(xnn_f32_vunary_ukernel_function vunary, OpType op_type, Variant variant = Variant::Native) const { |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 123 | std::random_device random_device; |
| 124 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 125 | auto distribution = std::uniform_real_distribution<float>(-125.0f, 125.0f); |
Marat Dukhan | f4db2f3 | 2020-06-30 10:55:30 -0700 | [diff] [blame] | 126 | auto f32rng = std::bind(distribution, std::ref(rng)); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 127 | |
| 128 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 129 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 130 | std::vector<double> y_ref(batch_size()); |
| 131 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 132 | if (inplace()) { |
| 133 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 134 | } else { |
| 135 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 136 | std::fill(y.begin(), y.end(), nanf("")); |
| 137 | } |
| 138 | const float* x_data = inplace() ? y.data() : x.data(); |
| 139 | |
| 140 | // Compute reference results. |
| 141 | for (size_t i = 0; i < batch_size(); i++) { |
| 142 | switch (op_type) { |
Frank Barchard | fb158e2 | 2020-07-15 16:10:10 -0700 | [diff] [blame] | 143 | case OpType::ReLU: |
| 144 | y_ref[i] = std::max(x_data[i], 0.0f); |
| 145 | break; |
Marat Dukhan | 0e80137 | 2022-01-04 00:10:41 -0800 | [diff] [blame] | 146 | default: |
| 147 | GTEST_FAIL() << "Unexpected operation type"; |
| 148 | return; |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 149 | } |
| 150 | } |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 151 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 152 | // Call optimized micro-kernel. |
Marat Dukhan | 0e80137 | 2022-01-04 00:10:41 -0800 | [diff] [blame] | 153 | vunary(batch_size() * sizeof(float), x_data, y.data(), nullptr); |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 154 | |
| 155 | // Verify results. |
| 156 | for (size_t i = 0; i < batch_size(); i++) { |
Frank Barchard | 2b9d29b | 2020-09-17 12:03:39 -0700 | [diff] [blame] | 157 | ASSERT_NEAR(y[i], y_ref[i], std::max(5.0e-6, std::abs(y_ref[i]) * 1.0e-5)) |
Marat Dukhan | 8d3c07e | 2020-01-02 01:20:59 -0800 | [diff] [blame] | 158 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 159 | } |
| 160 | } |
| 161 | } |
| 162 | |
Marat Dukhan | e5efb16 | 2021-12-31 10:26:13 -0800 | [diff] [blame] | 163 | void Test(xnn_f32_vabs_ukernel_function vabs, xnn_init_f32_abs_params_fn init_params = nullptr) const { |
| 164 | std::random_device random_device; |
| 165 | auto rng = std::mt19937(random_device()); |
| 166 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), std::ref(rng)); |
| 167 | |
| 168 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 169 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 170 | std::vector<float> y_ref(batch_size()); |
| 171 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 172 | if (inplace()) { |
| 173 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 174 | } else { |
| 175 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 176 | std::fill(y.begin(), y.end(), nanf("")); |
| 177 | } |
| 178 | const float* x_data = inplace() ? y.data() : x.data(); |
| 179 | |
| 180 | // Compute reference results. |
| 181 | for (size_t i = 0; i < batch_size(); i++) { |
| 182 | y_ref[i] = std::abs(x_data[i]); |
| 183 | } |
| 184 | |
| 185 | // Prepare parameters. |
| 186 | union xnn_f32_abs_params params; |
| 187 | if (init_params != nullptr) { |
| 188 | init_params(¶ms); |
| 189 | } |
| 190 | |
| 191 | // Call optimized micro-kernel. |
| 192 | vabs(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 193 | |
| 194 | // Verify results. |
| 195 | for (size_t i = 0; i < batch_size(); i++) { |
| 196 | ASSERT_EQ(y[i], y_ref[i]) |
| 197 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 198 | } |
| 199 | } |
| 200 | } |
| 201 | |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 202 | void Test(xnn_f32_vclamp_ukernel_function vclamp, xnn_init_f32_minmax_params_fn init_params) const { |
Marat Dukhan | 9491279 | 2021-08-16 21:40:30 -0700 | [diff] [blame] | 203 | std::random_device random_device; |
| 204 | auto rng = std::mt19937(random_device()); |
| 205 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 255.0f), std::ref(rng)); |
| 206 | |
| 207 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 208 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 209 | std::vector<float> y_ref(batch_size()); |
| 210 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 211 | if (inplace()) { |
| 212 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 213 | } else { |
| 214 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 215 | std::fill(y.begin(), y.end(), nanf("")); |
| 216 | } |
| 217 | const float* x_data = inplace() ? y.data() : x.data(); |
| 218 | |
| 219 | // Compute reference results. |
| 220 | for (size_t i = 0; i < batch_size(); i++) { |
| 221 | y_ref[i] = std::max(std::min(x_data[i], float(qmax())), float(qmin())); |
| 222 | } |
| 223 | |
| 224 | // Prepare parameters. |
| 225 | union xnn_f32_minmax_params params; |
| 226 | init_params(¶ms, float(qmin()), float(qmax())); |
| 227 | |
| 228 | // Call optimized micro-kernel. |
| 229 | vclamp(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 230 | |
| 231 | // Verify results. |
| 232 | for (size_t i = 0; i < batch_size(); i++) { |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 233 | ASSERT_EQ(y[i], y_ref[i]) |
Marat Dukhan | 9491279 | 2021-08-16 21:40:30 -0700 | [diff] [blame] | 234 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 235 | } |
| 236 | } |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 237 | } |
| 238 | |
Marat Dukhan | 4a79ff2 | 2022-01-01 12:16:48 -0800 | [diff] [blame] | 239 | void Test(xnn_f32_velu_ukernel_function velu, xnn_init_f32_elu_params_fn init_params) const { |
| 240 | std::random_device random_device; |
| 241 | auto rng = std::mt19937(random_device()); |
| 242 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-20.0f, 20.0f), std::ref(rng)); |
| 243 | |
| 244 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 245 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 246 | std::vector<double> y_ref(batch_size()); |
| 247 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 248 | if (inplace()) { |
| 249 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 250 | } else { |
| 251 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 252 | std::fill(y.begin(), y.end(), nanf("")); |
| 253 | } |
| 254 | const float* x_data = inplace() ? y.data() : x.data(); |
| 255 | |
| 256 | // Compute reference results. |
| 257 | for (size_t i = 0; i < batch_size(); i++) { |
| 258 | y_ref[i] = std::signbit(x_data[i]) ? alpha() * std::expm1(double(x_data[i]) * prescale()) : double(x_data[i]) * beta(); |
| 259 | } |
| 260 | |
| 261 | // Prepare parameters. |
| 262 | union xnn_f32_elu_params params; |
| 263 | init_params(¶ms, prescale(), alpha(), beta()); |
| 264 | |
| 265 | // Call optimized micro-kernel. |
| 266 | velu(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 267 | |
| 268 | // Verify results. |
| 269 | for (size_t i = 0; i < batch_size(); i++) { |
| 270 | ASSERT_NEAR(y[i], y_ref[i], std::max(5.0e-6, std::abs(y_ref[i]) * 1.0e-5)) |
| 271 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 272 | } |
| 273 | } |
| 274 | } |
| 275 | |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 276 | void Test(xnn_f32_vhswish_ukernel_function vhswish, xnn_init_f32_hswish_params_fn init_params) const { |
| 277 | std::random_device random_device; |
| 278 | auto rng = std::mt19937(random_device()); |
| 279 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-4.0f, 4.0f), std::ref(rng)); |
| 280 | |
| 281 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 282 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 283 | std::vector<double> y_ref(batch_size()); |
| 284 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 285 | if (inplace()) { |
| 286 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 287 | } else { |
| 288 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 289 | std::fill(y.begin(), y.end(), nanf("")); |
| 290 | } |
| 291 | const float* x_data = inplace() ? y.data() : x.data(); |
| 292 | |
| 293 | // Compute reference results. |
| 294 | for (size_t i = 0; i < batch_size(); i++) { |
| 295 | y_ref[i] = (x_data[i] / 6.0f) * std::max(std::min(x_data[i] + 3.0f, 6.0f), 0.0f); |
| 296 | } |
| 297 | |
| 298 | // Prepare parameters. |
| 299 | union xnn_f32_hswish_params params; |
| 300 | init_params(¶ms); |
| 301 | |
| 302 | // Call optimized micro-kernel. |
| 303 | vhswish(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 304 | |
| 305 | // Verify results. |
| 306 | for (size_t i = 0; i < batch_size(); i++) { |
| 307 | ASSERT_NEAR(y[i], y_ref[i], std::max(5.0e-6, std::abs(y_ref[i]) * 1.0e-5)) |
| 308 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 309 | } |
| 310 | } |
| 311 | } |
| 312 | |
Marat Dukhan | 2894e99 | 2021-12-30 08:29:48 -0800 | [diff] [blame] | 313 | void Test(xnn_f32_vlrelu_ukernel_function vlrelu, xnn_init_f32_lrelu_params_fn init_params) const { |
| 314 | std::random_device random_device; |
| 315 | auto rng = std::mt19937(random_device()); |
| 316 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-125.0f, 125.0f), std::ref(rng)); |
| 317 | |
| 318 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 319 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 320 | std::vector<double> y_ref(batch_size()); |
| 321 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 322 | if (inplace()) { |
| 323 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 324 | } else { |
| 325 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 326 | std::fill(y.begin(), y.end(), nanf("")); |
| 327 | } |
| 328 | const float* x_data = inplace() ? y.data() : x.data(); |
| 329 | |
| 330 | // Compute reference results. |
| 331 | for (size_t i = 0; i < batch_size(); i++) { |
| 332 | y_ref[i] = std::signbit(x_data[i]) ? x_data[i] * slope() : x_data[i]; |
| 333 | } |
| 334 | |
| 335 | // Prepare parameters. |
| 336 | union xnn_f32_lrelu_params params; |
| 337 | init_params(¶ms, slope()); |
| 338 | |
| 339 | // Call optimized micro-kernel. |
| 340 | vlrelu(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 341 | |
| 342 | // Verify results. |
| 343 | for (size_t i = 0; i < batch_size(); i++) { |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 344 | ASSERT_EQ(y[i], y_ref[i]) |
| 345 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 346 | } |
| 347 | } |
| 348 | } |
| 349 | |
Marat Dukhan | e5efb16 | 2021-12-31 10:26:13 -0800 | [diff] [blame] | 350 | void Test(xnn_f32_vneg_ukernel_function vneg, xnn_init_f32_neg_params_fn init_params = nullptr) const { |
| 351 | std::random_device random_device; |
| 352 | auto rng = std::mt19937(random_device()); |
| 353 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), std::ref(rng)); |
| 354 | |
| 355 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 356 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 357 | std::vector<float> y_ref(batch_size()); |
| 358 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 359 | if (inplace()) { |
| 360 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 361 | } else { |
| 362 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 363 | std::fill(y.begin(), y.end(), nanf("")); |
| 364 | } |
| 365 | const float* x_data = inplace() ? y.data() : x.data(); |
| 366 | |
| 367 | // Compute reference results. |
| 368 | for (size_t i = 0; i < batch_size(); i++) { |
| 369 | y_ref[i] = -x_data[i]; |
| 370 | } |
| 371 | |
| 372 | // Prepare parameters. |
| 373 | union xnn_f32_neg_params params; |
| 374 | if (init_params != nullptr) { |
| 375 | init_params(¶ms); |
| 376 | } |
| 377 | |
| 378 | // Call optimized micro-kernel. |
| 379 | vneg(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 380 | |
| 381 | // Verify results. |
| 382 | for (size_t i = 0; i < batch_size(); i++) { |
| 383 | ASSERT_EQ(y[i], y_ref[i]) |
| 384 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 385 | } |
| 386 | } |
| 387 | } |
| 388 | |
Marat Dukhan | 0e80137 | 2022-01-04 00:10:41 -0800 | [diff] [blame] | 389 | void Test(xnn_f32_vround_ukernel_function vrnd, OpType op_type, xnn_init_f32_rnd_params_fn init_params = nullptr) const { |
| 390 | std::random_device random_device; |
| 391 | auto rng = std::mt19937(random_device()); |
| 392 | auto distribution = std::uniform_real_distribution<float>(-5.0f, 5.0f); |
| 393 | auto f32rng = std::bind(distribution, std::ref(rng)); |
| 394 | |
| 395 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 396 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 397 | std::vector<float> y_ref(batch_size()); |
| 398 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 399 | if (inplace()) { |
| 400 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 401 | } else { |
| 402 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 403 | std::fill(y.begin(), y.end(), nanf("")); |
| 404 | } |
| 405 | const float* x_data = inplace() ? y.data() : x.data(); |
| 406 | |
| 407 | // Compute reference results. |
| 408 | for (size_t i = 0; i < batch_size(); i++) { |
| 409 | switch (op_type) { |
| 410 | case OpType::RoundToNearestEven: |
| 411 | y_ref[i] = std::nearbyint(double(x_data[i])); |
| 412 | break; |
| 413 | case OpType::RoundTowardsZero: |
| 414 | y_ref[i] = std::trunc(double(x_data[i])); |
| 415 | break; |
| 416 | case OpType::RoundUp: |
| 417 | y_ref[i] = std::ceil(double(x_data[i])); |
| 418 | break; |
| 419 | case OpType::RoundDown: |
| 420 | y_ref[i] = std::floor(double(x_data[i])); |
| 421 | break; |
| 422 | default: |
| 423 | GTEST_FAIL() << "Unexpected operation type"; |
| 424 | return; |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | // Prepare parameters. |
| 429 | xnn_f32_rnd_params params; |
| 430 | if (init_params != nullptr) { |
| 431 | init_params(¶ms); |
| 432 | } |
| 433 | |
| 434 | // Call optimized micro-kernel. |
| 435 | vrnd(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 436 | |
| 437 | // Verify results. |
| 438 | for (size_t i = 0; i < batch_size(); i++) { |
| 439 | ASSERT_EQ(y[i], y_ref[i]) |
| 440 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 441 | } |
| 442 | } |
| 443 | } |
| 444 | |
Marat Dukhan | ce834ad | 2022-01-03 00:22:01 -0800 | [diff] [blame] | 445 | void Test(xnn_f32_vsigmoid_ukernel_function vsigmoid, xnn_init_f32_sigmoid_params_fn init_params) const { |
| 446 | std::random_device random_device; |
| 447 | auto rng = std::mt19937(random_device()); |
| 448 | auto distribution = std::uniform_real_distribution<float>(-125.0f, 125.0f); |
| 449 | auto f32rng = std::bind(distribution, std::ref(rng)); |
| 450 | |
| 451 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 452 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 453 | std::vector<double> y_ref(batch_size()); |
| 454 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 455 | if (inplace()) { |
| 456 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 457 | } else { |
| 458 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 459 | std::fill(y.begin(), y.end(), nanf("")); |
| 460 | } |
| 461 | const float* x_data = inplace() ? y.data() : x.data(); |
| 462 | |
| 463 | // Compute reference results. |
| 464 | for (size_t i = 0; i < batch_size(); i++) { |
| 465 | const double e = std::exp(double(x_data[i])); |
| 466 | y_ref[i] = e / (1.0 + e); |
| 467 | } |
| 468 | |
| 469 | // Prepare parameters. |
| 470 | union xnn_f32_sigmoid_params params; |
| 471 | init_params(¶ms); |
| 472 | |
| 473 | // Call optimized micro-kernel. |
| 474 | vsigmoid(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 475 | |
| 476 | // Verify results. |
| 477 | for (size_t i = 0; i < batch_size(); i++) { |
| 478 | ASSERT_NEAR(y[i], y_ref[i], std::max(5.0e-6, std::abs(y_ref[i]) * 1.0e-5)) |
| 479 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 480 | } |
| 481 | } |
| 482 | } |
| 483 | |
Marat Dukhan | e5efb16 | 2021-12-31 10:26:13 -0800 | [diff] [blame] | 484 | void Test(xnn_f32_vsqr_ukernel_function vsqr, xnn_init_f32_default_params_fn init_params = nullptr) const { |
| 485 | std::random_device random_device; |
| 486 | auto rng = std::mt19937(random_device()); |
| 487 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-10.0f, 10.0f), std::ref(rng)); |
| 488 | |
| 489 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 490 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 491 | std::vector<float> y_ref(batch_size()); |
| 492 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 493 | if (inplace()) { |
| 494 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 495 | } else { |
| 496 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 497 | std::fill(y.begin(), y.end(), nanf("")); |
| 498 | } |
| 499 | const float* x_data = inplace() ? y.data() : x.data(); |
| 500 | |
| 501 | // Compute reference results. |
| 502 | for (size_t i = 0; i < batch_size(); i++) { |
| 503 | y_ref[i] = x_data[i] * x_data[i]; |
| 504 | } |
| 505 | |
| 506 | // Prepare parameters. |
| 507 | union xnn_f32_default_params params; |
| 508 | if (init_params != nullptr) { |
| 509 | init_params(¶ms); |
| 510 | } |
| 511 | |
| 512 | // Call optimized micro-kernel. |
| 513 | vsqr(batch_size() * sizeof(float), x_data, y.data(), ¶ms); |
| 514 | |
| 515 | // Verify results. |
| 516 | for (size_t i = 0; i < batch_size(); i++) { |
| 517 | ASSERT_EQ(y[i], y_ref[i]) |
| 518 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 519 | } |
| 520 | } |
| 521 | } |
| 522 | |
Marat Dukhan | e72b282 | 2021-12-30 14:46:58 -0800 | [diff] [blame] | 523 | void Test(xnn_f32_vsqrt_ukernel_function vsqrt, xnn_init_f32_sqrt_params_fn init_params = nullptr) const { |
| 524 | std::random_device random_device; |
| 525 | auto rng = std::mt19937(random_device()); |
| 526 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 10.0f), std::ref(rng)); |
| 527 | |
| 528 | std::vector<float> x(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 529 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 530 | std::vector<float> y_ref(batch_size()); |
| 531 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 532 | if (inplace()) { |
| 533 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 534 | } else { |
| 535 | std::generate(x.begin(), x.end(), std::ref(f32rng)); |
| 536 | std::fill(y.begin(), y.end(), nanf("")); |
| 537 | } |
| 538 | const float* x_data = inplace() ? y.data() : x.data(); |
| 539 | |
| 540 | // Compute reference results. |
| 541 | for (size_t i = 0; i < batch_size(); i++) { |
| 542 | y_ref[i] = std::sqrt(x_data[i]); |
| 543 | } |
| 544 | |
| 545 | // Prepare parameters. |
| 546 | union xnn_f32_sqrt_params params; |
| 547 | if (init_params != nullptr) { |
| 548 | init_params(¶ms); |
| 549 | } |
| 550 | |
| 551 | // Call optimized micro-kernel. |
| 552 | vsqrt(batch_size() * sizeof(float), x_data, y.data(), init_params != nullptr ? ¶ms : nullptr); |
| 553 | |
| 554 | // Verify results. |
| 555 | for (size_t i = 0; i < batch_size(); i++) { |
| 556 | ASSERT_EQ(y[i], y_ref[i]) |
Marat Dukhan | 2894e99 | 2021-12-30 08:29:48 -0800 | [diff] [blame] | 557 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << x[i]; |
| 558 | } |
| 559 | } |
| 560 | } |
| 561 | |
Marat Dukhan | 9491279 | 2021-08-16 21:40:30 -0700 | [diff] [blame] | 562 | inline void Test(xnn_f32_vabs_ukernel_function vunary, OpType op_type, Variant variant = Variant::Native) const { |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 563 | Test(xnn_f32_vunary_ukernel_function(vunary), op_type, variant); |
| 564 | } |
| 565 | |
| 566 | inline void Test(xnn_f32_velu_ukernel_function vunary, OpType op_type, Variant variant = Variant::Native) const { |
| 567 | Test(xnn_f32_vunary_ukernel_function(vunary), op_type, variant); |
| 568 | } |
| 569 | |
Marat Dukhan | 6eaab71 | 2021-05-13 15:20:58 -0700 | [diff] [blame] | 570 | inline void Test(xnn_f32_vneg_ukernel_function vunary, OpType op_type, Variant variant = Variant::Native) const { |
| 571 | Test(xnn_f32_vunary_ukernel_function(vunary), op_type, variant); |
| 572 | } |
| 573 | |
| 574 | inline void Test(xnn_f32_vrelu_ukernel_function vunary, OpType op_type, Variant variant = Variant::Native) const { |
| 575 | Test(xnn_f32_vunary_ukernel_function(vunary), op_type, variant); |
| 576 | } |
| 577 | |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 578 | void Test(xnn_f16_vclamp_ukernel_function vclamp, xnn_init_f16_minmax_params_fn init_params) const { |
| 579 | std::random_device random_device; |
| 580 | auto rng = std::mt19937(random_device()); |
| 581 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 255.0f), std::ref(rng)); |
| 582 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 583 | |
| 584 | std::vector<uint16_t> x(batch_size() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 585 | std::vector<uint16_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint16_t) : 0)); |
| 586 | std::vector<float> y_ref(batch_size()); |
| 587 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 588 | std::generate(x.begin(), x.end(), std::ref(f16rng)); |
| 589 | if (inplace()) { |
| 590 | std::generate(y.begin(), y.end(), std::ref(f16rng)); |
| 591 | } else { |
| 592 | std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */); |
| 593 | } |
| 594 | const uint16_t* x_data = inplace() ? y.data() : x.data(); |
| 595 | |
| 596 | // Compute reference results. |
| 597 | for (size_t i = 0; i < batch_size(); i++) { |
| 598 | y_ref[i] = std::max(std::min(fp16_ieee_to_fp32_value(x_data[i]), float(qmax())), float(qmin())); |
| 599 | } |
| 600 | |
| 601 | // Prepare parameters. |
Marat Dukhan | 14dd8d0 | 2022-01-06 16:03:31 -0800 | [diff] [blame] | 602 | union xnn_f16_minmax_params params; |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 603 | init_params(¶ms, fp16_ieee_from_fp32_value(float(qmin())), fp16_ieee_from_fp32_value(float(qmax()))); |
| 604 | |
| 605 | // Call optimized micro-kernel. |
| 606 | vclamp(batch_size() * sizeof(uint16_t), x_data, y.data(), ¶ms); |
| 607 | |
| 608 | // Verify results. |
| 609 | for (size_t i = 0; i < batch_size(); i++) { |
| 610 | ASSERT_NEAR(y_ref[i], fp16_ieee_to_fp32_value(y[i]), std::max(1.0e-3f, std::abs(y_ref[i]) * 1.0e-2f)) |
| 611 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << fp16_ieee_to_fp32_value(x[i]); |
| 612 | } |
| 613 | } |
Marat Dukhan | a6c0516 | 2021-05-13 16:52:02 -0700 | [diff] [blame] | 614 | } |
| 615 | |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 616 | void Test(xnn_f16_vhswish_ukernel_function vhswish, xnn_init_f16_hswish_params_fn init_params) const { |
| 617 | std::random_device random_device; |
| 618 | auto rng = std::mt19937(random_device()); |
| 619 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-4.0f, 4.0f), std::ref(rng)); |
| 620 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 621 | |
| 622 | std::vector<uint16_t> x(batch_size() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 623 | std::vector<uint16_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint16_t) : 0)); |
| 624 | std::vector<float> y_ref(batch_size()); |
| 625 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 626 | std::generate(x.begin(), x.end(), std::ref(f16rng)); |
| 627 | if (inplace()) { |
| 628 | std::generate(y.begin(), y.end(), std::ref(f16rng)); |
| 629 | } else { |
| 630 | std::fill(y.begin(), y.end(), UINT16_C(0x7E00) /* NaN */); |
| 631 | } |
| 632 | const uint16_t* x_data = inplace() ? y.data() : x.data(); |
| 633 | |
| 634 | // Compute reference results. |
| 635 | for (size_t i = 0; i < batch_size(); i++) { |
| 636 | const float x_value = fp16_ieee_to_fp32_value(x_data[i]); |
| 637 | y_ref[i] = (x_value / 6.0f) * std::max(std::min(x_value + 3.0f, 6.0f), 0.0f); |
| 638 | } |
| 639 | |
| 640 | // Prepare parameters. |
Marat Dukhan | 14dd8d0 | 2022-01-06 16:03:31 -0800 | [diff] [blame] | 641 | union xnn_f16_hswish_params params; |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 642 | init_params(¶ms); |
| 643 | |
| 644 | // Call optimized micro-kernel. |
| 645 | vhswish(batch_size() * sizeof(uint16_t), x_data, y.data(), ¶ms); |
| 646 | |
| 647 | // Verify results. |
| 648 | for (size_t i = 0; i < batch_size(); i++) { |
| 649 | ASSERT_NEAR(y_ref[i], fp16_ieee_to_fp32_value(y[i]), std::max(1.0e-3f, std::abs(y_ref[i]) * 1.0e-2f)) |
| 650 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << fp16_ieee_to_fp32_value(x[i]); |
| 651 | } |
| 652 | } |
Marat Dukhan | a6c0516 | 2021-05-13 16:52:02 -0700 | [diff] [blame] | 653 | } |
| 654 | |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 655 | void Test(xnn_s8_vclamp_ukernel_function vclamp, xnn_init_s8_minmax_params_fn init_params) const { |
Marat Dukhan | e79acb7 | 2021-08-16 19:03:53 -0700 | [diff] [blame] | 656 | std::random_device random_device; |
| 657 | auto rng = std::mt19937(random_device()); |
| 658 | auto i8rng = std::bind( |
| 659 | std::uniform_int_distribution<int32_t>(std::numeric_limits<int8_t>::min(), std::numeric_limits<int8_t>::max()), |
| 660 | std::ref(rng)); |
| 661 | |
| 662 | std::vector<int8_t> x(batch_size() + XNN_EXTRA_BYTES / sizeof(int8_t)); |
| 663 | std::vector<int8_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(int8_t) : 0)); |
| 664 | std::vector<int8_t> y_ref(batch_size()); |
| 665 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 666 | std::generate(x.begin(), x.end(), std::ref(i8rng)); |
| 667 | if (inplace()) { |
| 668 | std::copy(x.cbegin(), x.cend(), y.begin()); |
| 669 | } else { |
| 670 | std::fill(y.begin(), y.end(), INT8_C(0xA5)); |
| 671 | } |
| 672 | const int8_t* x_data = inplace() ? y.data() : x.data(); |
| 673 | |
| 674 | // Compute reference results. |
| 675 | for (size_t i = 0; i < batch_size(); i++) { |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 676 | y_ref[i] = std::min(std::max(x_data[i], int8_t(qmin() - 0x80)), int8_t(qmax() - 0x80)); |
Marat Dukhan | e79acb7 | 2021-08-16 19:03:53 -0700 | [diff] [blame] | 677 | } |
| 678 | |
| 679 | // Prepare parameters. |
| 680 | union xnn_s8_minmax_params params; |
| 681 | init_params(¶ms, int8_t(qmin() - 0x80), int8_t(qmax() - 0x80)); |
| 682 | |
| 683 | // Call optimized micro-kernel. |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 684 | vclamp(batch_size() * sizeof(int8_t), x_data, y.data(), ¶ms); |
Marat Dukhan | e79acb7 | 2021-08-16 19:03:53 -0700 | [diff] [blame] | 685 | |
| 686 | // Verify results. |
| 687 | for (size_t i = 0; i < batch_size(); i++) { |
| 688 | ASSERT_EQ(int32_t(y_ref[i]), int32_t(y[i])) |
| 689 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << int32_t(x[i]); |
| 690 | } |
| 691 | } |
| 692 | } |
| 693 | |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 694 | void Test(xnn_u8_vclamp_ukernel_function vclamp, xnn_init_u8_minmax_params_fn init_params) const { |
Marat Dukhan | a6c0516 | 2021-05-13 16:52:02 -0700 | [diff] [blame] | 695 | std::random_device random_device; |
| 696 | auto rng = std::mt19937(random_device()); |
Marat Dukhan | 1f5b108 | 2021-08-16 17:01:44 -0700 | [diff] [blame] | 697 | auto u8rng = std::bind( |
| 698 | std::uniform_int_distribution<int32_t>(0, std::numeric_limits<uint8_t>::max()), std::ref(rng)); |
Marat Dukhan | a6c0516 | 2021-05-13 16:52:02 -0700 | [diff] [blame] | 699 | |
| 700 | std::vector<uint8_t> x(batch_size() + XNN_EXTRA_BYTES / sizeof(uint8_t)); |
| 701 | std::vector<uint8_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint8_t) : 0)); |
| 702 | std::vector<uint8_t> y_ref(batch_size()); |
| 703 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 704 | std::generate(x.begin(), x.end(), std::ref(u8rng)); |
| 705 | if (inplace()) { |
| 706 | std::copy(x.cbegin(), x.cend(), y.begin()); |
| 707 | } else { |
| 708 | std::fill(y.begin(), y.end(), UINT8_C(0xA5)); |
| 709 | } |
| 710 | const uint8_t* x_data = inplace() ? y.data() : x.data(); |
| 711 | |
| 712 | // Compute reference results. |
| 713 | for (size_t i = 0; i < batch_size(); i++) { |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 714 | y_ref[i] = std::min(std::max(x_data[i], qmin()), qmax()); |
Marat Dukhan | a6c0516 | 2021-05-13 16:52:02 -0700 | [diff] [blame] | 715 | } |
| 716 | |
| 717 | // Prepare parameters. |
Marat Dukhan | 1f5b108 | 2021-08-16 17:01:44 -0700 | [diff] [blame] | 718 | union xnn_u8_minmax_params params; |
| 719 | init_params(¶ms, qmin(), qmax()); |
Marat Dukhan | a6c0516 | 2021-05-13 16:52:02 -0700 | [diff] [blame] | 720 | |
| 721 | // Call optimized micro-kernel. |
Marat Dukhan | 0d10cc7 | 2021-12-23 19:49:19 -0800 | [diff] [blame] | 722 | vclamp(batch_size() * sizeof(uint8_t), x_data, y.data(), ¶ms); |
Marat Dukhan | a6c0516 | 2021-05-13 16:52:02 -0700 | [diff] [blame] | 723 | |
| 724 | // Verify results. |
| 725 | for (size_t i = 0; i < batch_size(); i++) { |
| 726 | ASSERT_EQ(uint32_t(y_ref[i]), uint32_t(y[i])) |
| 727 | << "at " << i << " / " << batch_size() << ", x[" << i << "] = " << uint32_t(x[i]); |
| 728 | } |
| 729 | } |
| 730 | } |
| 731 | |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 732 | private: |
Marat Dukhan | ed6baaf | 2020-12-01 15:07:08 -0800 | [diff] [blame] | 733 | size_t batch_size_ = 1; |
| 734 | bool inplace_ = false; |
| 735 | float slope_ = 0.5f; |
| 736 | float prescale_ = 1.0f; |
| 737 | float alpha_ = 1.0f; |
| 738 | float beta_ = 1.0f; |
| 739 | uint8_t qmin_ = 0; |
| 740 | uint8_t qmax_ = 255; |
| 741 | size_t iterations_ = 15; |
Marat Dukhan | 346a9e5 | 2019-11-15 09:06:30 -0800 | [diff] [blame] | 742 | }; |