Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -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 | |
| 18 | #include <xnnpack.h> |
| 19 | #include <xnnpack/params-init.h> |
| 20 | #include <xnnpack/params.h> |
| 21 | |
| 22 | |
| 23 | class VBinOpCMicrokernelTester { |
| 24 | public: |
| 25 | enum class OpType { |
| 26 | AddC, |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 27 | DivC, |
| 28 | RDivC, |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 29 | MaxC, |
| 30 | MinC, |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 31 | MulC, |
| 32 | SubC, |
| 33 | RSubC, |
| 34 | }; |
| 35 | |
| 36 | enum class Variant { |
| 37 | Native, |
| 38 | Scalar, |
| 39 | }; |
| 40 | |
| 41 | inline VBinOpCMicrokernelTester& batch_size(size_t batch_size) { |
| 42 | assert(batch_size != 0); |
| 43 | this->batch_size_ = batch_size; |
| 44 | return *this; |
| 45 | } |
| 46 | |
| 47 | inline size_t batch_size() const { |
| 48 | return this->batch_size_; |
| 49 | } |
| 50 | |
| 51 | inline VBinOpCMicrokernelTester& inplace(bool inplace) { |
| 52 | this->inplace_ = inplace; |
| 53 | return *this; |
| 54 | } |
| 55 | |
| 56 | inline bool inplace() const { |
| 57 | return this->inplace_; |
| 58 | } |
| 59 | |
| 60 | inline VBinOpCMicrokernelTester& qmin(uint8_t qmin) { |
| 61 | this->qmin_ = qmin; |
| 62 | return *this; |
| 63 | } |
| 64 | |
| 65 | inline uint8_t qmin() const { |
| 66 | return this->qmin_; |
| 67 | } |
| 68 | |
| 69 | inline VBinOpCMicrokernelTester& qmax(uint8_t qmax) { |
| 70 | this->qmax_ = qmax; |
| 71 | return *this; |
| 72 | } |
| 73 | |
| 74 | inline uint8_t qmax() const { |
| 75 | return this->qmax_; |
| 76 | } |
| 77 | |
| 78 | inline VBinOpCMicrokernelTester& iterations(size_t iterations) { |
| 79 | this->iterations_ = iterations; |
| 80 | return *this; |
| 81 | } |
| 82 | |
| 83 | inline size_t iterations() const { |
| 84 | return this->iterations_; |
| 85 | } |
| 86 | |
Frank Barchard | bf31e3f | 2020-05-12 14:00:07 -0700 | [diff] [blame] | 87 | void Test(xnn_f16_vbinary_ukernel_function vbinaryc, OpType op_type) const { |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 88 | std::random_device random_device; |
| 89 | auto rng = std::mt19937(random_device()); |
| 90 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.01f, 1.0f), rng); |
| 91 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 92 | |
| 93 | std::vector<uint16_t> a(batch_size() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 94 | const uint16_t b = f16rng(); |
| 95 | std::vector<uint16_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint16_t) : 0)); |
| 96 | std::vector<float> y_ref(batch_size()); |
| 97 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 98 | std::generate(a.begin(), a.end(), std::ref(f16rng)); |
| 99 | if (inplace()) { |
| 100 | std::generate(y.begin(), y.end(), std::ref(f16rng)); |
| 101 | } else { |
| 102 | std::fill(y.begin(), y.end(), nanf("")); |
| 103 | } |
| 104 | const uint16_t* a_data = inplace() ? y.data() : a.data(); |
| 105 | |
| 106 | // Compute reference results. |
| 107 | for (size_t i = 0; i < batch_size(); i++) { |
| 108 | switch (op_type) { |
| 109 | case OpType::AddC: |
| 110 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) + fp16_ieee_to_fp32_value(b); |
| 111 | break; |
| 112 | case OpType::DivC: |
| 113 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) / fp16_ieee_to_fp32_value(b); |
| 114 | break; |
| 115 | case OpType::RDivC: |
| 116 | y_ref[i] = fp16_ieee_to_fp32_value(b) / fp16_ieee_to_fp32_value(a_data[i]); |
| 117 | break; |
| 118 | case OpType::MaxC: |
| 119 | y_ref[i] = std::max<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 120 | break; |
| 121 | case OpType::MinC: |
| 122 | y_ref[i] = std::min<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 123 | break; |
| 124 | case OpType::MulC: |
| 125 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) * fp16_ieee_to_fp32_value(b); |
| 126 | break; |
| 127 | case OpType::SubC: |
| 128 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) - fp16_ieee_to_fp32_value(b); |
| 129 | break; |
| 130 | case OpType::RSubC: |
| 131 | y_ref[i] = fp16_ieee_to_fp32_value(b) - fp16_ieee_to_fp32_value(a_data[i]); |
| 132 | break; |
| 133 | } |
| 134 | } |
| 135 | // Call optimized micro-kernel. |
| 136 | vbinaryc(batch_size() * sizeof(uint16_t), a_data, &b, y.data(), nullptr); |
| 137 | |
| 138 | // Verify results. |
| 139 | for (size_t i = 0; i < batch_size(); i++) { |
| 140 | ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::abs(y_ref[i]) * 1.0e-2f) |
| 141 | << "at " << i << " / " << batch_size(); |
| 142 | } |
| 143 | } |
| 144 | } |
| 145 | |
Frank Barchard | bf31e3f | 2020-05-12 14:00:07 -0700 | [diff] [blame] | 146 | void Test(xnn_f16_vbinary_minmax_ukernel_function vbinaryc_minmax, OpType op_type) const { |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 147 | std::random_device random_device; |
| 148 | auto rng = std::mt19937(random_device()); |
| 149 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.01f, 1.0f), rng); |
| 150 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 151 | |
| 152 | std::vector<uint16_t> a(batch_size() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 153 | const uint16_t b = f16rng(); |
| 154 | std::vector<uint16_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint16_t) : 0)); |
| 155 | std::vector<float> y_ref(batch_size()); |
| 156 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 157 | std::generate(a.begin(), a.end(), std::ref(f16rng)); |
| 158 | if (inplace()) { |
| 159 | std::generate(y.begin(), y.end(), std::ref(f16rng)); |
| 160 | } else { |
| 161 | std::fill(y.begin(), y.end(), nanf("")); |
| 162 | } |
| 163 | const uint16_t* a_data = inplace() ? y.data() : a.data(); |
| 164 | |
| 165 | // Compute reference results. |
| 166 | for (size_t i = 0; i < batch_size(); i++) { |
| 167 | switch (op_type) { |
| 168 | case OpType::AddC: |
| 169 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) + fp16_ieee_to_fp32_value(b); |
| 170 | break; |
| 171 | case OpType::DivC: |
| 172 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) / fp16_ieee_to_fp32_value(b); |
| 173 | break; |
| 174 | case OpType::RDivC: |
| 175 | y_ref[i] = fp16_ieee_to_fp32_value(b) / fp16_ieee_to_fp32_value(a_data[i]); |
| 176 | break; |
| 177 | case OpType::MaxC: |
| 178 | y_ref[i] = std::max<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 179 | break; |
| 180 | case OpType::MinC: |
| 181 | y_ref[i] = std::min<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 182 | break; |
| 183 | case OpType::MulC: |
| 184 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) * fp16_ieee_to_fp32_value(b); |
| 185 | break; |
| 186 | case OpType::SubC: |
| 187 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) - fp16_ieee_to_fp32_value(b); |
| 188 | break; |
| 189 | case OpType::RSubC: |
| 190 | y_ref[i] = fp16_ieee_to_fp32_value(b) - fp16_ieee_to_fp32_value(a_data[i]); |
| 191 | break; |
| 192 | } |
| 193 | } |
| 194 | const float accumulated_min = *std::min_element(y_ref.cbegin(), y_ref.cend()); |
| 195 | const float accumulated_max = *std::max_element(y_ref.cbegin(), y_ref.cend()); |
| 196 | const float accumulated_range = accumulated_max - accumulated_min; |
| 197 | const float y_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_range > 0.0f ? |
| 198 | (accumulated_max - accumulated_range / 255.0f * float(255 - qmax())) : |
| 199 | +std::numeric_limits<float>::infinity())); |
| 200 | const float y_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_range > 0.0f ? |
| 201 | (accumulated_min + accumulated_range / 255.0f * float(qmin())) : |
| 202 | -std::numeric_limits<float>::infinity())); |
| 203 | for (size_t i = 0; i < batch_size(); i++) { |
| 204 | y_ref[i] = std::max<float>(std::min<float>(y_ref[i], y_max), y_min); |
| 205 | } |
| 206 | |
Frank Barchard | 9f3a843 | 2020-06-02 13:59:35 -0700 | [diff] [blame^] | 207 | // Prepare parameters. |
Frank Barchard | bf31e3f | 2020-05-12 14:00:07 -0700 | [diff] [blame] | 208 | xnn_f16_minmax_params params = xnn_init_f16_minmax_params( |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 209 | fp16_ieee_from_fp32_value(y_min), |
| 210 | fp16_ieee_from_fp32_value(y_max)); |
| 211 | |
| 212 | // Call optimized micro-kernel. |
| 213 | vbinaryc_minmax(batch_size() * sizeof(uint16_t), a_data, &b, y.data(), ¶ms); |
| 214 | |
| 215 | // Verify results. |
| 216 | for (size_t i = 0; i < batch_size(); i++) { |
| 217 | ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::abs(y_ref[i]) * 1.0e-2f) |
| 218 | << "at " << i << " / " << batch_size(); |
| 219 | } |
| 220 | } |
| 221 | } |
| 222 | |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 223 | void Test(xnn_f32_vbinary_ukernel_function vbinaryc, OpType op_type, Variant variant = Variant::Native) const { |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 224 | std::random_device random_device; |
| 225 | auto rng = std::mt19937(random_device()); |
| 226 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 227 | |
| 228 | std::vector<float> a(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 229 | const float b = f32rng(); |
| 230 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 231 | std::vector<float> y_ref(batch_size()); |
| 232 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 233 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 234 | if (inplace()) { |
| 235 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 236 | } else { |
| 237 | std::fill(y.begin(), y.end(), nanf("")); |
| 238 | } |
| 239 | const float* a_data = inplace() ? y.data() : a.data(); |
| 240 | |
| 241 | // Compute reference results. |
| 242 | for (size_t i = 0; i < batch_size(); i++) { |
| 243 | switch (op_type) { |
| 244 | case OpType::AddC: |
| 245 | y_ref[i] = a_data[i] + b; |
| 246 | break; |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 247 | case OpType::DivC: |
| 248 | y_ref[i] = a_data[i] / b; |
| 249 | break; |
| 250 | case OpType::RDivC: |
| 251 | y_ref[i] = b / a_data[i]; |
| 252 | break; |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 253 | case OpType::MaxC: |
| 254 | y_ref[i] = std::max<float>(a_data[i], b); |
| 255 | break; |
| 256 | case OpType::MinC: |
| 257 | y_ref[i] = std::min<float>(a_data[i], b); |
| 258 | break; |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 259 | case OpType::MulC: |
| 260 | y_ref[i] = a_data[i] * b; |
| 261 | break; |
| 262 | case OpType::SubC: |
| 263 | y_ref[i] = a_data[i] - b; |
| 264 | break; |
| 265 | case OpType::RSubC: |
| 266 | y_ref[i] = b - a_data[i]; |
| 267 | break; |
| 268 | } |
| 269 | } |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 270 | // Call optimized micro-kernel. |
| 271 | vbinaryc(batch_size() * sizeof(float), a_data, &b, y.data(), nullptr); |
| 272 | |
| 273 | // Verify results. |
| 274 | for (size_t i = 0; i < batch_size(); i++) { |
| 275 | ASSERT_NEAR(y[i], y_ref[i], std::abs(y_ref[i]) * 1.0e-6f) |
| 276 | << "at " << i << " / " << batch_size(); |
| 277 | } |
| 278 | } |
| 279 | } |
| 280 | |
| 281 | void Test(xnn_f32_vbinary_minmax_ukernel_function vbinaryc_minmax, OpType op_type, Variant variant = Variant::Native) const { |
| 282 | std::random_device random_device; |
| 283 | auto rng = std::mt19937(random_device()); |
| 284 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 285 | |
| 286 | std::vector<float> a(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 287 | const float b = f32rng(); |
| 288 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 289 | std::vector<float> y_ref(batch_size()); |
| 290 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 291 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 292 | if (inplace()) { |
| 293 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 294 | } else { |
| 295 | std::fill(y.begin(), y.end(), nanf("")); |
| 296 | } |
| 297 | const float* a_data = inplace() ? y.data() : a.data(); |
| 298 | |
| 299 | // Compute reference results. |
| 300 | for (size_t i = 0; i < batch_size(); i++) { |
| 301 | switch (op_type) { |
| 302 | case OpType::AddC: |
| 303 | y_ref[i] = a_data[i] + b; |
| 304 | break; |
| 305 | case OpType::DivC: |
| 306 | y_ref[i] = a_data[i] / b; |
| 307 | break; |
| 308 | case OpType::RDivC: |
| 309 | y_ref[i] = b / a_data[i]; |
| 310 | break; |
| 311 | case OpType::MaxC: |
| 312 | y_ref[i] = std::max<float>(a_data[i], b); |
| 313 | break; |
| 314 | case OpType::MinC: |
| 315 | y_ref[i] = std::min<float>(a_data[i], b); |
| 316 | break; |
| 317 | case OpType::MulC: |
| 318 | y_ref[i] = a_data[i] * b; |
| 319 | break; |
| 320 | case OpType::SubC: |
| 321 | y_ref[i] = a_data[i] - b; |
| 322 | break; |
| 323 | case OpType::RSubC: |
| 324 | y_ref[i] = b - a_data[i]; |
| 325 | break; |
| 326 | } |
| 327 | } |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 328 | const float accumulated_min = *std::min_element(y_ref.cbegin(), y_ref.cend()); |
| 329 | const float accumulated_max = *std::max_element(y_ref.cbegin(), y_ref.cend()); |
| 330 | const float accumulated_range = accumulated_max - accumulated_min; |
| 331 | const float y_max = accumulated_range > 0.0f ? |
| 332 | (accumulated_max - accumulated_range / 255.0f * float(255 - qmax())) : |
| 333 | +std::numeric_limits<float>::infinity(); |
| 334 | const float y_min = accumulated_range > 0.0f ? |
| 335 | (accumulated_min + accumulated_range / 255.0f * float(qmin())) : |
| 336 | -std::numeric_limits<float>::infinity(); |
| 337 | for (size_t i = 0; i < batch_size(); i++) { |
| 338 | y_ref[i] = std::max<float>(std::min<float>(y_ref[i], y_max), y_min); |
| 339 | } |
| 340 | |
Frank Barchard | 9f3a843 | 2020-06-02 13:59:35 -0700 | [diff] [blame^] | 341 | // Prepare parameters. |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 342 | xnn_f32_minmax_params params = { }; |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 343 | switch (variant) { |
| 344 | case Variant::Native: |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 345 | params = xnn_init_f32_minmax_params(y_min, y_max); |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 346 | break; |
| 347 | case Variant::Scalar: |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 348 | params = xnn_init_scalar_f32_minmax_params(y_min, y_max); |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 349 | break; |
| 350 | } |
| 351 | |
| 352 | // Call optimized micro-kernel. |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 353 | vbinaryc_minmax(batch_size() * sizeof(float), a_data, &b, y.data(), ¶ms); |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 354 | |
| 355 | // Verify results. |
| 356 | for (size_t i = 0; i < batch_size(); i++) { |
| 357 | ASSERT_NEAR(y[i], y_ref[i], std::abs(y_ref[i]) * 1.0e-6f) |
| 358 | << "at " << i << " / " << batch_size(); |
| 359 | } |
| 360 | } |
| 361 | } |
| 362 | |
| 363 | private: |
| 364 | size_t batch_size_{1}; |
| 365 | bool inplace_{false}; |
| 366 | uint8_t qmin_{0}; |
| 367 | uint8_t qmax_{255}; |
| 368 | size_t iterations_{15}; |
| 369 | }; |