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 | |
Frank Barchard | 9c1a735 | 2020-06-04 20:15:01 -0700 | [diff] [blame] | 18 | #include <fp16.h> |
| 19 | |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 20 | #include <xnnpack.h> |
| 21 | #include <xnnpack/params-init.h> |
| 22 | #include <xnnpack/params.h> |
| 23 | |
| 24 | |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 25 | class VBinaryCMicrokernelTester { |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 26 | public: |
| 27 | enum class OpType { |
| 28 | AddC, |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 29 | DivC, |
| 30 | RDivC, |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 31 | MaxC, |
| 32 | MinC, |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 33 | MulC, |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 34 | SqrDiffC, |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 35 | SubC, |
| 36 | RSubC, |
| 37 | }; |
| 38 | |
| 39 | enum class Variant { |
| 40 | Native, |
| 41 | Scalar, |
| 42 | }; |
| 43 | |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 44 | inline VBinaryCMicrokernelTester& batch_size(size_t batch_size) { |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 45 | assert(batch_size != 0); |
| 46 | this->batch_size_ = batch_size; |
| 47 | return *this; |
| 48 | } |
| 49 | |
| 50 | inline size_t batch_size() const { |
| 51 | return this->batch_size_; |
| 52 | } |
| 53 | |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 54 | inline VBinaryCMicrokernelTester& inplace(bool inplace) { |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 55 | this->inplace_ = inplace; |
| 56 | return *this; |
| 57 | } |
| 58 | |
| 59 | inline bool inplace() const { |
| 60 | return this->inplace_; |
| 61 | } |
| 62 | |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 63 | inline VBinaryCMicrokernelTester& qmin(uint8_t qmin) { |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 64 | this->qmin_ = qmin; |
| 65 | return *this; |
| 66 | } |
| 67 | |
| 68 | inline uint8_t qmin() const { |
| 69 | return this->qmin_; |
| 70 | } |
| 71 | |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 72 | inline VBinaryCMicrokernelTester& qmax(uint8_t qmax) { |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 73 | this->qmax_ = qmax; |
| 74 | return *this; |
| 75 | } |
| 76 | |
| 77 | inline uint8_t qmax() const { |
| 78 | return this->qmax_; |
| 79 | } |
| 80 | |
Marat Dukhan | 10f1fe0 | 2021-05-13 12:51:12 -0700 | [diff] [blame] | 81 | inline VBinaryCMicrokernelTester& iterations(size_t iterations) { |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 82 | this->iterations_ = iterations; |
| 83 | return *this; |
| 84 | } |
| 85 | |
| 86 | inline size_t iterations() const { |
| 87 | return this->iterations_; |
| 88 | } |
| 89 | |
Frank Barchard | bf31e3f | 2020-05-12 14:00:07 -0700 | [diff] [blame] | 90 | void Test(xnn_f16_vbinary_ukernel_function vbinaryc, OpType op_type) const { |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 91 | std::random_device random_device; |
| 92 | auto rng = std::mt19937(random_device()); |
Frank Barchard | 967712d | 2021-03-22 12:01:44 -0700 | [diff] [blame] | 93 | auto f32rng = std::bind(std::uniform_real_distribution<float>(1.0e-3f, 1.0f), rng); |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 94 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 95 | |
| 96 | std::vector<uint16_t> a(batch_size() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 97 | const uint16_t b = f16rng(); |
| 98 | std::vector<uint16_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint16_t) : 0)); |
| 99 | std::vector<float> y_ref(batch_size()); |
| 100 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 101 | std::generate(a.begin(), a.end(), std::ref(f16rng)); |
| 102 | if (inplace()) { |
| 103 | std::generate(y.begin(), y.end(), std::ref(f16rng)); |
| 104 | } else { |
| 105 | std::fill(y.begin(), y.end(), nanf("")); |
| 106 | } |
| 107 | const uint16_t* a_data = inplace() ? y.data() : a.data(); |
| 108 | |
| 109 | // Compute reference results. |
| 110 | for (size_t i = 0; i < batch_size(); i++) { |
| 111 | switch (op_type) { |
| 112 | case OpType::AddC: |
| 113 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) + fp16_ieee_to_fp32_value(b); |
| 114 | break; |
| 115 | case OpType::DivC: |
| 116 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) / fp16_ieee_to_fp32_value(b); |
| 117 | break; |
| 118 | case OpType::RDivC: |
| 119 | y_ref[i] = fp16_ieee_to_fp32_value(b) / fp16_ieee_to_fp32_value(a_data[i]); |
| 120 | break; |
| 121 | case OpType::MaxC: |
| 122 | y_ref[i] = std::max<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 123 | break; |
| 124 | case OpType::MinC: |
| 125 | y_ref[i] = std::min<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 126 | break; |
| 127 | case OpType::MulC: |
| 128 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) * fp16_ieee_to_fp32_value(b); |
| 129 | break; |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 130 | case OpType::SqrDiffC: |
| 131 | { |
| 132 | const float diff = fp16_ieee_to_fp32_value(a_data[i]) - fp16_ieee_to_fp32_value(b); |
| 133 | y_ref[i] = diff * diff; |
| 134 | break; |
| 135 | } |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 136 | case OpType::SubC: |
| 137 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) - fp16_ieee_to_fp32_value(b); |
| 138 | break; |
| 139 | case OpType::RSubC: |
| 140 | y_ref[i] = fp16_ieee_to_fp32_value(b) - fp16_ieee_to_fp32_value(a_data[i]); |
| 141 | break; |
| 142 | } |
| 143 | } |
| 144 | // Call optimized micro-kernel. |
| 145 | vbinaryc(batch_size() * sizeof(uint16_t), a_data, &b, y.data(), nullptr); |
| 146 | |
| 147 | // Verify results. |
| 148 | for (size_t i = 0; i < batch_size(); i++) { |
Frank Barchard | 2b9d29b | 2020-09-17 12:03:39 -0700 | [diff] [blame] | 149 | ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-4f, std::abs(y_ref[i]) * 1.0e-2f)) |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 150 | << "at " << i << " / " << batch_size(); |
| 151 | } |
| 152 | } |
| 153 | } |
| 154 | |
Frank Barchard | bf31e3f | 2020-05-12 14:00:07 -0700 | [diff] [blame] | 155 | 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] | 156 | std::random_device random_device; |
| 157 | auto rng = std::mt19937(random_device()); |
Frank Barchard | 967712d | 2021-03-22 12:01:44 -0700 | [diff] [blame] | 158 | auto f32rng = std::bind(std::uniform_real_distribution<float>(1.0e-3f, 1.0f), rng); |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 159 | auto f16rng = std::bind(fp16_ieee_from_fp32_value, f32rng); |
| 160 | |
| 161 | std::vector<uint16_t> a(batch_size() + XNN_EXTRA_BYTES / sizeof(uint16_t)); |
| 162 | const uint16_t b = f16rng(); |
| 163 | std::vector<uint16_t> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(uint16_t) : 0)); |
| 164 | std::vector<float> y_ref(batch_size()); |
| 165 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 166 | std::generate(a.begin(), a.end(), std::ref(f16rng)); |
| 167 | if (inplace()) { |
| 168 | std::generate(y.begin(), y.end(), std::ref(f16rng)); |
| 169 | } else { |
| 170 | std::fill(y.begin(), y.end(), nanf("")); |
| 171 | } |
| 172 | const uint16_t* a_data = inplace() ? y.data() : a.data(); |
| 173 | |
| 174 | // Compute reference results. |
| 175 | for (size_t i = 0; i < batch_size(); i++) { |
| 176 | switch (op_type) { |
| 177 | case OpType::AddC: |
| 178 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) + fp16_ieee_to_fp32_value(b); |
| 179 | break; |
| 180 | case OpType::DivC: |
| 181 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) / fp16_ieee_to_fp32_value(b); |
| 182 | break; |
| 183 | case OpType::RDivC: |
| 184 | y_ref[i] = fp16_ieee_to_fp32_value(b) / fp16_ieee_to_fp32_value(a_data[i]); |
| 185 | break; |
| 186 | case OpType::MaxC: |
| 187 | y_ref[i] = std::max<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 188 | break; |
| 189 | case OpType::MinC: |
| 190 | y_ref[i] = std::min<float>(fp16_ieee_to_fp32_value(a_data[i]), fp16_ieee_to_fp32_value(b)); |
| 191 | break; |
| 192 | case OpType::MulC: |
| 193 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) * fp16_ieee_to_fp32_value(b); |
| 194 | break; |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 195 | case OpType::SqrDiffC: |
| 196 | { |
| 197 | const float diff = fp16_ieee_to_fp32_value(a_data[i]) - fp16_ieee_to_fp32_value(b); |
| 198 | y_ref[i] = diff * diff; |
| 199 | break; |
| 200 | } |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 201 | case OpType::SubC: |
| 202 | y_ref[i] = fp16_ieee_to_fp32_value(a_data[i]) - fp16_ieee_to_fp32_value(b); |
| 203 | break; |
| 204 | case OpType::RSubC: |
| 205 | y_ref[i] = fp16_ieee_to_fp32_value(b) - fp16_ieee_to_fp32_value(a_data[i]); |
| 206 | break; |
| 207 | } |
| 208 | } |
| 209 | const float accumulated_min = *std::min_element(y_ref.cbegin(), y_ref.cend()); |
| 210 | const float accumulated_max = *std::max_element(y_ref.cbegin(), y_ref.cend()); |
| 211 | const float accumulated_range = accumulated_max - accumulated_min; |
| 212 | const float y_max = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_range > 0.0f ? |
| 213 | (accumulated_max - accumulated_range / 255.0f * float(255 - qmax())) : |
| 214 | +std::numeric_limits<float>::infinity())); |
| 215 | const float y_min = fp16_ieee_to_fp32_value(fp16_ieee_from_fp32_value(accumulated_range > 0.0f ? |
| 216 | (accumulated_min + accumulated_range / 255.0f * float(qmin())) : |
| 217 | -std::numeric_limits<float>::infinity())); |
| 218 | for (size_t i = 0; i < batch_size(); i++) { |
| 219 | y_ref[i] = std::max<float>(std::min<float>(y_ref[i], y_max), y_min); |
| 220 | } |
| 221 | |
Frank Barchard | 9f3a843 | 2020-06-02 13:59:35 -0700 | [diff] [blame] | 222 | // Prepare parameters. |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame^] | 223 | xnn_f16_minmax_params params; |
| 224 | xnn_init_f16_minmax_params( |
| 225 | ¶ms, |
| 226 | fp16_ieee_from_fp32_value(y_min), |
| 227 | fp16_ieee_from_fp32_value(y_max)); |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 228 | |
| 229 | // Call optimized micro-kernel. |
| 230 | vbinaryc_minmax(batch_size() * sizeof(uint16_t), a_data, &b, y.data(), ¶ms); |
| 231 | |
| 232 | // Verify results. |
| 233 | for (size_t i = 0; i < batch_size(); i++) { |
Frank Barchard | 2b9d29b | 2020-09-17 12:03:39 -0700 | [diff] [blame] | 234 | ASSERT_NEAR(fp16_ieee_to_fp32_value(y[i]), y_ref[i], std::max(1.0e-4f, std::abs(y_ref[i]) * 1.0e-2f)) |
Frank Barchard | d793f6c | 2020-05-08 13:37:43 -0700 | [diff] [blame] | 235 | << "at " << i << " / " << batch_size(); |
| 236 | } |
| 237 | } |
| 238 | } |
| 239 | |
Marat Dukhan | 1e782c4 | 2019-11-21 17:02:40 -0800 | [diff] [blame] | 240 | 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] | 241 | std::random_device random_device; |
| 242 | auto rng = std::mt19937(random_device()); |
| 243 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 244 | |
| 245 | std::vector<float> a(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 246 | const float b = f32rng(); |
| 247 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 248 | std::vector<float> y_ref(batch_size()); |
| 249 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 250 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 251 | if (inplace()) { |
| 252 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 253 | } else { |
| 254 | std::fill(y.begin(), y.end(), nanf("")); |
| 255 | } |
| 256 | const float* a_data = inplace() ? y.data() : a.data(); |
| 257 | |
| 258 | // Compute reference results. |
| 259 | for (size_t i = 0; i < batch_size(); i++) { |
| 260 | switch (op_type) { |
| 261 | case OpType::AddC: |
| 262 | y_ref[i] = a_data[i] + b; |
| 263 | break; |
Marat Dukhan | 77ca630 | 2019-12-06 12:48:15 -0800 | [diff] [blame] | 264 | case OpType::DivC: |
| 265 | y_ref[i] = a_data[i] / b; |
| 266 | break; |
| 267 | case OpType::RDivC: |
| 268 | y_ref[i] = b / a_data[i]; |
| 269 | break; |
Marat Dukhan | 403b7d4 | 2019-12-05 12:49:11 -0800 | [diff] [blame] | 270 | case OpType::MaxC: |
| 271 | y_ref[i] = std::max<float>(a_data[i], b); |
| 272 | break; |
| 273 | case OpType::MinC: |
| 274 | y_ref[i] = std::min<float>(a_data[i], b); |
| 275 | break; |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 276 | case OpType::MulC: |
| 277 | y_ref[i] = a_data[i] * b; |
| 278 | break; |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 279 | case OpType::SqrDiffC: |
| 280 | { |
| 281 | const float diff = a_data[i] - b; |
| 282 | y_ref[i] = diff * diff; |
| 283 | break; |
| 284 | } |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 285 | case OpType::SubC: |
| 286 | y_ref[i] = a_data[i] - b; |
| 287 | break; |
| 288 | case OpType::RSubC: |
| 289 | y_ref[i] = b - a_data[i]; |
| 290 | break; |
| 291 | } |
| 292 | } |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 293 | // Call optimized micro-kernel. |
| 294 | vbinaryc(batch_size() * sizeof(float), a_data, &b, y.data(), nullptr); |
| 295 | |
| 296 | // Verify results. |
| 297 | for (size_t i = 0; i < batch_size(); i++) { |
| 298 | ASSERT_NEAR(y[i], y_ref[i], std::abs(y_ref[i]) * 1.0e-6f) |
| 299 | << "at " << i << " / " << batch_size(); |
| 300 | } |
| 301 | } |
| 302 | } |
| 303 | |
| 304 | void Test(xnn_f32_vbinary_minmax_ukernel_function vbinaryc_minmax, OpType op_type, Variant variant = Variant::Native) const { |
| 305 | std::random_device random_device; |
| 306 | auto rng = std::mt19937(random_device()); |
| 307 | auto f32rng = std::bind(std::uniform_real_distribution<float>(0.0f, 1.0f), rng); |
| 308 | |
| 309 | std::vector<float> a(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 310 | const float b = f32rng(); |
| 311 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 312 | std::vector<float> y_ref(batch_size()); |
| 313 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 314 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 315 | if (inplace()) { |
| 316 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 317 | } else { |
| 318 | std::fill(y.begin(), y.end(), nanf("")); |
| 319 | } |
| 320 | const float* a_data = inplace() ? y.data() : a.data(); |
| 321 | |
| 322 | // Compute reference results. |
| 323 | for (size_t i = 0; i < batch_size(); i++) { |
| 324 | switch (op_type) { |
| 325 | case OpType::AddC: |
| 326 | y_ref[i] = a_data[i] + b; |
| 327 | break; |
| 328 | case OpType::DivC: |
| 329 | y_ref[i] = a_data[i] / b; |
| 330 | break; |
| 331 | case OpType::RDivC: |
| 332 | y_ref[i] = b / a_data[i]; |
| 333 | break; |
| 334 | case OpType::MaxC: |
| 335 | y_ref[i] = std::max<float>(a_data[i], b); |
| 336 | break; |
| 337 | case OpType::MinC: |
| 338 | y_ref[i] = std::min<float>(a_data[i], b); |
| 339 | break; |
| 340 | case OpType::MulC: |
| 341 | y_ref[i] = a_data[i] * b; |
| 342 | break; |
Marat Dukhan | 13bafb0 | 2020-06-05 00:43:11 -0700 | [diff] [blame] | 343 | case OpType::SqrDiffC: |
| 344 | { |
| 345 | const float diff = a_data[i] - b; |
| 346 | y_ref[i] = diff * diff; |
| 347 | break; |
| 348 | } |
Marat Dukhan | 91cd2b7 | 2020-04-09 23:57:31 -0700 | [diff] [blame] | 349 | case OpType::SubC: |
| 350 | y_ref[i] = a_data[i] - b; |
| 351 | break; |
| 352 | case OpType::RSubC: |
| 353 | y_ref[i] = b - a_data[i]; |
| 354 | break; |
| 355 | } |
| 356 | } |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 357 | const float accumulated_min = *std::min_element(y_ref.cbegin(), y_ref.cend()); |
| 358 | const float accumulated_max = *std::max_element(y_ref.cbegin(), y_ref.cend()); |
| 359 | const float accumulated_range = accumulated_max - accumulated_min; |
| 360 | const float y_max = accumulated_range > 0.0f ? |
| 361 | (accumulated_max - accumulated_range / 255.0f * float(255 - qmax())) : |
| 362 | +std::numeric_limits<float>::infinity(); |
| 363 | const float y_min = accumulated_range > 0.0f ? |
| 364 | (accumulated_min + accumulated_range / 255.0f * float(qmin())) : |
| 365 | -std::numeric_limits<float>::infinity(); |
| 366 | for (size_t i = 0; i < batch_size(); i++) { |
| 367 | y_ref[i] = std::max<float>(std::min<float>(y_ref[i], y_max), y_min); |
| 368 | } |
| 369 | |
Frank Barchard | 9f3a843 | 2020-06-02 13:59:35 -0700 | [diff] [blame] | 370 | // Prepare parameters. |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame^] | 371 | xnn_f32_minmax_params params; |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 372 | switch (variant) { |
| 373 | case Variant::Native: |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame^] | 374 | xnn_init_f32_minmax_params(¶ms, y_min, y_max); |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 375 | break; |
| 376 | case Variant::Scalar: |
Marat Dukhan | f56f4c4 | 2021-05-17 01:47:20 -0700 | [diff] [blame^] | 377 | xnn_init_scalar_f32_minmax_params(¶ms, y_min, y_max); |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 378 | break; |
| 379 | } |
| 380 | |
| 381 | // Call optimized micro-kernel. |
Frank Barchard | e70dbeb | 2020-05-01 15:46:41 -0700 | [diff] [blame] | 382 | vbinaryc_minmax(batch_size() * sizeof(float), a_data, &b, y.data(), ¶ms); |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 383 | |
| 384 | // Verify results. |
| 385 | for (size_t i = 0; i < batch_size(); i++) { |
| 386 | ASSERT_NEAR(y[i], y_ref[i], std::abs(y_ref[i]) * 1.0e-6f) |
| 387 | << "at " << i << " / " << batch_size(); |
| 388 | } |
| 389 | } |
| 390 | } |
| 391 | |
Frank Barchard | 674778d | 2020-08-08 10:17:25 -0700 | [diff] [blame] | 392 | void Test(xnn_f32_vbinary_relu_ukernel_function vbinaryc_relu, OpType op_type, Variant variant = Variant::Native) const { |
| 393 | std::random_device random_device; |
| 394 | auto rng = std::mt19937(random_device()); |
| 395 | auto f32rng = std::bind(std::uniform_real_distribution<float>(-1.0f, 1.0f), rng); |
| 396 | |
| 397 | std::vector<float> a(batch_size() + XNN_EXTRA_BYTES / sizeof(float)); |
| 398 | const float b = f32rng(); |
| 399 | std::vector<float> y(batch_size() + (inplace() ? XNN_EXTRA_BYTES / sizeof(float) : 0)); |
| 400 | std::vector<float> y_ref(batch_size()); |
| 401 | for (size_t iteration = 0; iteration < iterations(); iteration++) { |
| 402 | std::generate(a.begin(), a.end(), std::ref(f32rng)); |
| 403 | if (inplace()) { |
| 404 | std::generate(y.begin(), y.end(), std::ref(f32rng)); |
| 405 | } else { |
| 406 | std::fill(y.begin(), y.end(), nanf("")); |
| 407 | } |
| 408 | const float* a_data = inplace() ? y.data() : a.data(); |
| 409 | |
| 410 | // Compute reference results. |
| 411 | for (size_t i = 0; i < batch_size(); i++) { |
| 412 | switch (op_type) { |
| 413 | case OpType::AddC: |
| 414 | y_ref[i] = a_data[i] + b; |
| 415 | break; |
| 416 | case OpType::DivC: |
| 417 | y_ref[i] = a_data[i] / b; |
| 418 | break; |
| 419 | case OpType::RDivC: |
| 420 | y_ref[i] = b / a_data[i]; |
| 421 | break; |
| 422 | case OpType::MaxC: |
| 423 | y_ref[i] = std::max<float>(a_data[i], b); |
| 424 | break; |
| 425 | case OpType::MinC: |
| 426 | y_ref[i] = std::min<float>(a_data[i], b); |
| 427 | break; |
| 428 | case OpType::MulC: |
| 429 | y_ref[i] = a_data[i] * b; |
| 430 | break; |
| 431 | case OpType::SqrDiffC: |
| 432 | { |
| 433 | const float diff = a_data[i] - b; |
| 434 | y_ref[i] = diff * diff; |
| 435 | break; |
| 436 | } |
| 437 | case OpType::SubC: |
| 438 | y_ref[i] = a_data[i] - b; |
| 439 | break; |
| 440 | case OpType::RSubC: |
| 441 | y_ref[i] = b - a_data[i]; |
| 442 | break; |
| 443 | } |
| 444 | } |
| 445 | for (size_t i = 0; i < batch_size(); i++) { |
| 446 | y_ref[i] = std::max(y_ref[i], 0.0f); |
| 447 | } |
| 448 | |
| 449 | // Prepare parameters. |
| 450 | xnn_f32_relu_params params = { }; |
| 451 | |
| 452 | // Call optimized micro-kernel. |
| 453 | vbinaryc_relu(batch_size() * sizeof(float), a_data, &b, y.data(), ¶ms); |
| 454 | |
| 455 | // Verify results. |
| 456 | for (size_t i = 0; i < batch_size(); i++) { |
| 457 | ASSERT_GE(y[i], 0.0f) |
| 458 | << "at " << i << " / " << batch_size(); |
| 459 | ASSERT_NEAR(y[i], y_ref[i], std::abs(y_ref[i]) * 1.0e-6f) |
| 460 | << "at " << i << " / " << batch_size(); |
| 461 | } |
| 462 | } |
| 463 | } |
| 464 | |
Marat Dukhan | c07cb7f | 2019-11-14 15:32:05 -0800 | [diff] [blame] | 465 | private: |
| 466 | size_t batch_size_{1}; |
| 467 | bool inplace_{false}; |
| 468 | uint8_t qmin_{0}; |
| 469 | uint8_t qmax_{255}; |
| 470 | size_t iterations_{15}; |
| 471 | }; |