Marat Dukhan | 329da64 | 2019-11-19 21:44:39 -0800 | [diff] [blame] | 1 | #!/usr/bin/env python |
| 2 | # Copyright 2019 Google LLC |
| 3 | # |
| 4 | # This source code is licensed under the BSD-style license found in the |
| 5 | # LICENSE file in the root directory of this source tree. |
| 6 | |
| 7 | import argparse |
| 8 | import codecs |
| 9 | import math |
| 10 | import os |
| 11 | import re |
| 12 | import sys |
| 13 | import yaml |
| 14 | |
| 15 | sys.path.insert(0, os.path.dirname(os.path.abspath(__file__))) |
| 16 | from primes import next_prime |
| 17 | import xngen |
| 18 | import xnncommon |
| 19 | |
| 20 | |
| 21 | parser = argparse.ArgumentParser(description='MaxPool microkernel test generator') |
| 22 | parser.add_argument("-s", "--spec", metavar="FILE", required=True, |
| 23 | help="Specification (YAML) file") |
| 24 | parser.add_argument("-o", "--output", metavar="FILE", required=True, |
| 25 | help='Output (C++ source) file') |
| 26 | parser.set_defaults(defines=list()) |
| 27 | |
| 28 | |
| 29 | def split_ukernel_name(name): |
| 30 | match = re.match(r"^xnn_(s8|u8|s16|f16|f32)_maxpool_ukernel_(\d+)p(\d+)x__(.+)_c(\d+)$", name) |
| 31 | if match is None: |
| 32 | raise ValueError("Unexpected microkernel name: " + name) |
| 33 | |
| 34 | primary_tile = int(match.group(2)) |
| 35 | incremental_tile = int(match.group(3)) |
| 36 | channel_tile = int(match.group(5)) |
| 37 | |
| 38 | arch, isa = xnncommon.parse_target_name(target_name=match.group(4)) |
| 39 | return primary_tile, incremental_tile, channel_tile, arch, isa |
| 40 | |
| 41 | |
| 42 | MAXPOOL_TEST_TEMPLATE = """\ |
| 43 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile) { |
| 44 | $if ISA_CHECK: |
| 45 | ${ISA_CHECK}; |
| 46 | MaxPoolMicrokernelTester() |
| 47 | .pooling_elements(${PRIMARY_TILE}) |
| 48 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 49 | .channels(${CHANNEL_TILE}) |
| 50 | .Test(${", ".join(TEST_ARGS)}); |
| 51 | } |
| 52 | |
| 53 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 54 | $if ISA_CHECK: |
| 55 | ${ISA_CHECK}; |
| 56 | MaxPoolMicrokernelTester() |
| 57 | .pooling_elements(${PRIMARY_TILE}) |
| 58 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 59 | .channels(${CHANNEL_TILE}) |
| 60 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 61 | .Test(${", ".join(TEST_ARGS)}); |
| 62 | } |
| 63 | |
| 64 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 65 | $if ISA_CHECK: |
| 66 | ${ISA_CHECK}; |
| 67 | MaxPoolMicrokernelTester() |
| 68 | .pooling_elements(${PRIMARY_TILE}) |
| 69 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 70 | .channels(${CHANNEL_TILE}) |
| 71 | .qmin(192) |
| 72 | .Test(${", ".join(TEST_ARGS)}); |
| 73 | } |
| 74 | |
| 75 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 76 | $if ISA_CHECK: |
| 77 | ${ISA_CHECK}; |
| 78 | MaxPoolMicrokernelTester() |
| 79 | .pooling_elements(${PRIMARY_TILE}) |
| 80 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 81 | .channels(${CHANNEL_TILE}) |
| 82 | .qmax(192) |
| 83 | .Test(${", ".join(TEST_ARGS)}); |
| 84 | } |
| 85 | |
| 86 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_subtile) { |
| 87 | $if ISA_CHECK: |
| 88 | ${ISA_CHECK}; |
| 89 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 90 | MaxPoolMicrokernelTester() |
| 91 | .pooling_elements(pooling_elements) |
| 92 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 93 | .channels(${CHANNEL_TILE}) |
| 94 | .Test(${", ".join(TEST_ARGS)}); |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 99 | $if ISA_CHECK: |
| 100 | ${ISA_CHECK}; |
| 101 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 102 | MaxPoolMicrokernelTester() |
| 103 | .pooling_elements(pooling_elements) |
| 104 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 105 | .channels(${CHANNEL_TILE}) |
| 106 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 107 | .Test(${", ".join(TEST_ARGS)}); |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | $if CHANNEL_TILE > 1: |
| 112 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile) { |
| 113 | $if ISA_CHECK: |
| 114 | ${ISA_CHECK}; |
| 115 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 116 | MaxPoolMicrokernelTester() |
| 117 | .pooling_elements(${PRIMARY_TILE}) |
| 118 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 119 | .channels(channels) |
| 120 | .Test(${", ".join(TEST_ARGS)}); |
| 121 | } |
| 122 | } |
| 123 | |
| 124 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 125 | $if ISA_CHECK: |
| 126 | ${ISA_CHECK}; |
| 127 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 128 | MaxPoolMicrokernelTester() |
| 129 | .pooling_elements(${PRIMARY_TILE}) |
| 130 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 131 | .channels(channels) |
| 132 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 133 | .Test(${", ".join(TEST_ARGS)}); |
| 134 | } |
| 135 | } |
| 136 | |
| 137 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 138 | $if ISA_CHECK: |
| 139 | ${ISA_CHECK}; |
| 140 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 141 | MaxPoolMicrokernelTester() |
| 142 | .pooling_elements(${PRIMARY_TILE}) |
| 143 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 144 | .channels(channels) |
| 145 | .qmin(192) |
| 146 | .Test(${", ".join(TEST_ARGS)}); |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 151 | $if ISA_CHECK: |
| 152 | ${ISA_CHECK}; |
| 153 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 154 | MaxPoolMicrokernelTester() |
| 155 | .pooling_elements(${PRIMARY_TILE}) |
| 156 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 157 | .channels(channels) |
| 158 | .qmax(192) |
| 159 | .Test(${", ".join(TEST_ARGS)}); |
| 160 | } |
| 161 | } |
| 162 | |
| 163 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_subtile) { |
| 164 | $if ISA_CHECK: |
| 165 | ${ISA_CHECK}; |
| 166 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 167 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 168 | MaxPoolMicrokernelTester() |
| 169 | .pooling_elements(pooling_elements) |
| 170 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 171 | .channels(channels) |
| 172 | .Test(${", ".join(TEST_ARGS)}); |
| 173 | } |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 178 | $if ISA_CHECK: |
| 179 | ${ISA_CHECK}; |
| 180 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 181 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 182 | MaxPoolMicrokernelTester() |
| 183 | .pooling_elements(pooling_elements) |
| 184 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 185 | .channels(channels) |
| 186 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 187 | .Test(${", ".join(TEST_ARGS)}); |
| 188 | } |
| 189 | } |
| 190 | } |
| 191 | |
| 192 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile) { |
| 193 | $if ISA_CHECK: |
| 194 | ${ISA_CHECK}; |
| 195 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 196 | MaxPoolMicrokernelTester() |
| 197 | .pooling_elements(${PRIMARY_TILE}) |
| 198 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 199 | .channels(channels) |
| 200 | .Test(${", ".join(TEST_ARGS)}); |
| 201 | } |
| 202 | } |
| 203 | |
| 204 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 205 | $if ISA_CHECK: |
| 206 | ${ISA_CHECK}; |
| 207 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 208 | MaxPoolMicrokernelTester() |
| 209 | .pooling_elements(${PRIMARY_TILE}) |
| 210 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 211 | .channels(channels) |
| 212 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 213 | .Test(${", ".join(TEST_ARGS)}); |
| 214 | } |
| 215 | } |
| 216 | |
| 217 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 218 | $if ISA_CHECK: |
| 219 | ${ISA_CHECK}; |
| 220 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 221 | MaxPoolMicrokernelTester() |
| 222 | .pooling_elements(${PRIMARY_TILE}) |
| 223 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 224 | .channels(channels) |
| 225 | .qmin(192) |
| 226 | .Test(${", ".join(TEST_ARGS)}); |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 231 | $if ISA_CHECK: |
| 232 | ${ISA_CHECK}; |
| 233 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 234 | MaxPoolMicrokernelTester() |
| 235 | .pooling_elements(${PRIMARY_TILE}) |
| 236 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 237 | .channels(channels) |
| 238 | .qmax(192) |
| 239 | .Test(${", ".join(TEST_ARGS)}); |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_subtile) { |
| 244 | $if ISA_CHECK: |
| 245 | ${ISA_CHECK}; |
| 246 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 247 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 248 | MaxPoolMicrokernelTester() |
| 249 | .pooling_elements(pooling_elements) |
| 250 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 251 | .channels(channels) |
| 252 | .Test(${", ".join(TEST_ARGS)}); |
| 253 | } |
| 254 | } |
| 255 | } |
| 256 | |
| 257 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 258 | $if ISA_CHECK: |
| 259 | ${ISA_CHECK}; |
| 260 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 261 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 262 | MaxPoolMicrokernelTester() |
| 263 | .pooling_elements(pooling_elements) |
| 264 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 265 | .channels(channels) |
| 266 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 267 | .Test(${", ".join(TEST_ARGS)}); |
| 268 | } |
| 269 | } |
| 270 | } |
| 271 | |
| 272 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile) { |
| 273 | $if ISA_CHECK: |
| 274 | ${ISA_CHECK}; |
| 275 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 276 | MaxPoolMicrokernelTester() |
| 277 | .pooling_elements(${PRIMARY_TILE}) |
| 278 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 279 | .channels(channels) |
| 280 | .Test(${", ".join(TEST_ARGS)}); |
| 281 | } |
| 282 | } |
| 283 | |
| 284 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 285 | $if ISA_CHECK: |
| 286 | ${ISA_CHECK}; |
| 287 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 288 | MaxPoolMicrokernelTester() |
| 289 | .pooling_elements(${PRIMARY_TILE}) |
| 290 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 291 | .channels(channels) |
| 292 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 293 | .Test(${", ".join(TEST_ARGS)}); |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 298 | $if ISA_CHECK: |
| 299 | ${ISA_CHECK}; |
| 300 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 301 | MaxPoolMicrokernelTester() |
| 302 | .pooling_elements(${PRIMARY_TILE}) |
| 303 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 304 | .channels(channels) |
| 305 | .qmin(192) |
| 306 | .Test(${", ".join(TEST_ARGS)}); |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 311 | $if ISA_CHECK: |
| 312 | ${ISA_CHECK}; |
| 313 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 314 | MaxPoolMicrokernelTester() |
| 315 | .pooling_elements(${PRIMARY_TILE}) |
| 316 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 317 | .channels(channels) |
| 318 | .qmax(192) |
| 319 | .Test(${", ".join(TEST_ARGS)}); |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_subtile) { |
| 324 | $if ISA_CHECK: |
| 325 | ${ISA_CHECK}; |
| 326 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 327 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 328 | MaxPoolMicrokernelTester() |
| 329 | .pooling_elements(pooling_elements) |
| 330 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 331 | .channels(channels) |
| 332 | .Test(${", ".join(TEST_ARGS)}); |
| 333 | } |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 338 | $if ISA_CHECK: |
| 339 | ${ISA_CHECK}; |
| 340 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 341 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 342 | MaxPoolMicrokernelTester() |
| 343 | .pooling_elements(pooling_elements) |
| 344 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 345 | .channels(channels) |
| 346 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 347 | .Test(${", ".join(TEST_ARGS)}); |
| 348 | } |
| 349 | } |
| 350 | } |
| 351 | |
| 352 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile) { |
| 353 | $if ISA_CHECK: |
| 354 | ${ISA_CHECK}; |
| 355 | MaxPoolMicrokernelTester() |
| 356 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 357 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 358 | .channels(${CHANNEL_TILE}) |
| 359 | .Test(${", ".join(TEST_ARGS)}); |
| 360 | } |
| 361 | |
| 362 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 363 | $if ISA_CHECK: |
| 364 | ${ISA_CHECK}; |
| 365 | MaxPoolMicrokernelTester() |
| 366 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 367 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 368 | .channels(${CHANNEL_TILE}) |
| 369 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 370 | .Test(${", ".join(TEST_ARGS)}); |
| 371 | } |
| 372 | |
| 373 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 374 | $if ISA_CHECK: |
| 375 | ${ISA_CHECK}; |
| 376 | MaxPoolMicrokernelTester() |
| 377 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 378 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 379 | .channels(${CHANNEL_TILE}) |
| 380 | .qmin(192) |
| 381 | .Test(${", ".join(TEST_ARGS)}); |
| 382 | } |
| 383 | |
| 384 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 385 | $if ISA_CHECK: |
| 386 | ${ISA_CHECK}; |
| 387 | MaxPoolMicrokernelTester() |
| 388 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 389 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 390 | .channels(${CHANNEL_TILE}) |
| 391 | .qmax(192) |
| 392 | .Test(${", ".join(TEST_ARGS)}); |
| 393 | } |
| 394 | |
| 395 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_subtile) { |
| 396 | $if ISA_CHECK: |
| 397 | ${ISA_CHECK}; |
| 398 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 399 | MaxPoolMicrokernelTester() |
| 400 | .pooling_elements(pooling_elements) |
| 401 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 402 | .channels(${CHANNEL_TILE}) |
| 403 | .Test(${", ".join(TEST_ARGS)}); |
| 404 | } |
| 405 | } |
| 406 | |
| 407 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 408 | $if ISA_CHECK: |
| 409 | ${ISA_CHECK}; |
| 410 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 411 | MaxPoolMicrokernelTester() |
| 412 | .pooling_elements(pooling_elements) |
| 413 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 414 | .channels(${CHANNEL_TILE}) |
| 415 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 416 | .Test(${", ".join(TEST_ARGS)}); |
| 417 | } |
| 418 | } |
| 419 | |
| 420 | $if CHANNEL_TILE > 1: |
| 421 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile) { |
| 422 | $if ISA_CHECK: |
| 423 | ${ISA_CHECK}; |
| 424 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 425 | MaxPoolMicrokernelTester() |
| 426 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 427 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 428 | .channels(channels) |
| 429 | .Test(${", ".join(TEST_ARGS)}); |
| 430 | } |
| 431 | } |
| 432 | |
| 433 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 434 | $if ISA_CHECK: |
| 435 | ${ISA_CHECK}; |
| 436 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 437 | MaxPoolMicrokernelTester() |
| 438 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 439 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 440 | .channels(channels) |
| 441 | .input_offset(${next_prime(CHANNEL_TILE*5)}) |
| 442 | .Test(${", ".join(TEST_ARGS)}); |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 447 | $if ISA_CHECK: |
| 448 | ${ISA_CHECK}; |
| 449 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 450 | MaxPoolMicrokernelTester() |
| 451 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 452 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 453 | .channels(channels) |
| 454 | .qmin(192) |
| 455 | .Test(${", ".join(TEST_ARGS)}); |
| 456 | } |
| 457 | } |
| 458 | |
| 459 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 460 | $if ISA_CHECK: |
| 461 | ${ISA_CHECK}; |
| 462 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 463 | MaxPoolMicrokernelTester() |
| 464 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 465 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 466 | .channels(channels) |
| 467 | .qmax(192) |
| 468 | .Test(${", ".join(TEST_ARGS)}); |
| 469 | } |
| 470 | } |
| 471 | |
| 472 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_subtile) { |
| 473 | $if ISA_CHECK: |
| 474 | ${ISA_CHECK}; |
| 475 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 476 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 477 | MaxPoolMicrokernelTester() |
| 478 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 479 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 480 | .channels(channels) |
| 481 | .Test(${", ".join(TEST_ARGS)}); |
| 482 | } |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 487 | $if ISA_CHECK: |
| 488 | ${ISA_CHECK}; |
| 489 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 490 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 491 | MaxPoolMicrokernelTester() |
| 492 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 493 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 494 | .channels(channels) |
| 495 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 496 | .Test(${", ".join(TEST_ARGS)}); |
| 497 | } |
| 498 | } |
| 499 | } |
| 500 | |
| 501 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile) { |
| 502 | $if ISA_CHECK: |
| 503 | ${ISA_CHECK}; |
| 504 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 505 | MaxPoolMicrokernelTester() |
| 506 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 507 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 508 | .channels(channels) |
| 509 | .Test(${", ".join(TEST_ARGS)}); |
| 510 | } |
| 511 | } |
| 512 | |
| 513 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 514 | $if ISA_CHECK: |
| 515 | ${ISA_CHECK}; |
| 516 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 517 | MaxPoolMicrokernelTester() |
| 518 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 519 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 520 | .channels(channels) |
| 521 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 522 | .Test(${", ".join(TEST_ARGS)}); |
| 523 | } |
| 524 | } |
| 525 | |
| 526 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 527 | $if ISA_CHECK: |
| 528 | ${ISA_CHECK}; |
| 529 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 530 | MaxPoolMicrokernelTester() |
| 531 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 532 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 533 | .channels(channels) |
| 534 | .qmin(192) |
| 535 | .Test(${", ".join(TEST_ARGS)}); |
| 536 | } |
| 537 | } |
| 538 | |
| 539 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 540 | $if ISA_CHECK: |
| 541 | ${ISA_CHECK}; |
| 542 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 543 | MaxPoolMicrokernelTester() |
| 544 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 545 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 546 | .channels(channels) |
| 547 | .qmax(192) |
| 548 | .Test(${", ".join(TEST_ARGS)}); |
| 549 | } |
| 550 | } |
| 551 | |
| 552 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_subtile) { |
| 553 | $if ISA_CHECK: |
| 554 | ${ISA_CHECK}; |
| 555 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 556 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 557 | MaxPoolMicrokernelTester() |
| 558 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 559 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 560 | .channels(channels) |
| 561 | .Test(${", ".join(TEST_ARGS)}); |
| 562 | } |
| 563 | } |
| 564 | } |
| 565 | |
| 566 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 567 | $if ISA_CHECK: |
| 568 | ${ISA_CHECK}; |
| 569 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 570 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 571 | MaxPoolMicrokernelTester() |
| 572 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 573 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 574 | .channels(channels) |
| 575 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 576 | .Test(${", ".join(TEST_ARGS)}); |
| 577 | } |
| 578 | } |
| 579 | } |
| 580 | |
| 581 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile) { |
| 582 | $if ISA_CHECK: |
| 583 | ${ISA_CHECK}; |
| 584 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 585 | MaxPoolMicrokernelTester() |
| 586 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 587 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 588 | .channels(channels) |
| 589 | .Test(${", ".join(TEST_ARGS)}); |
| 590 | } |
| 591 | } |
| 592 | |
| 593 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 594 | $if ISA_CHECK: |
| 595 | ${ISA_CHECK}; |
| 596 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 597 | MaxPoolMicrokernelTester() |
| 598 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 599 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 600 | .channels(channels) |
| 601 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 602 | .Test(${", ".join(TEST_ARGS)}); |
| 603 | } |
| 604 | } |
| 605 | |
| 606 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 607 | $if ISA_CHECK: |
| 608 | ${ISA_CHECK}; |
| 609 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 610 | MaxPoolMicrokernelTester() |
| 611 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 612 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 613 | .channels(channels) |
| 614 | .qmin(192) |
| 615 | .Test(${", ".join(TEST_ARGS)}); |
| 616 | } |
| 617 | } |
| 618 | |
| 619 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 620 | $if ISA_CHECK: |
| 621 | ${ISA_CHECK}; |
| 622 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 623 | MaxPoolMicrokernelTester() |
| 624 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 625 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 626 | .channels(channels) |
| 627 | .qmax(192) |
| 628 | .Test(${", ".join(TEST_ARGS)}); |
| 629 | } |
| 630 | } |
| 631 | |
| 632 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_subtile) { |
| 633 | $if ISA_CHECK: |
| 634 | ${ISA_CHECK}; |
| 635 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 636 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 637 | MaxPoolMicrokernelTester() |
| 638 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 639 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 640 | .channels(channels) |
| 641 | .Test(${", ".join(TEST_ARGS)}); |
| 642 | } |
| 643 | } |
| 644 | } |
| 645 | |
| 646 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 647 | $if ISA_CHECK: |
| 648 | ${ISA_CHECK}; |
| 649 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 650 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 651 | MaxPoolMicrokernelTester() |
| 652 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 653 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 654 | .channels(channels) |
| 655 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 656 | .Test(${", ".join(TEST_ARGS)}); |
| 657 | } |
| 658 | } |
| 659 | } |
| 660 | |
| 661 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass) { |
| 662 | $if ISA_CHECK: |
| 663 | ${ISA_CHECK}; |
| 664 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 665 | MaxPoolMicrokernelTester() |
| 666 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 667 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 668 | .channels(${CHANNEL_TILE}) |
| 669 | .Test(${", ".join(TEST_ARGS)}); |
| 670 | } |
| 671 | } |
| 672 | |
| 673 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 674 | $if ISA_CHECK: |
| 675 | ${ISA_CHECK}; |
| 676 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 677 | MaxPoolMicrokernelTester() |
| 678 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 679 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 680 | .channels(${CHANNEL_TILE}) |
| 681 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 682 | .Test(${", ".join(TEST_ARGS)}); |
| 683 | } |
| 684 | } |
| 685 | |
| 686 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass_with_qmin) { |
| 687 | $if ISA_CHECK: |
| 688 | ${ISA_CHECK}; |
| 689 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 690 | MaxPoolMicrokernelTester() |
| 691 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 692 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 693 | .channels(${CHANNEL_TILE}) |
| 694 | .qmin(192) |
| 695 | .Test(${", ".join(TEST_ARGS)}); |
| 696 | } |
| 697 | } |
| 698 | |
| 699 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass_with_qmax) { |
| 700 | $if ISA_CHECK: |
| 701 | ${ISA_CHECK}; |
| 702 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 703 | MaxPoolMicrokernelTester() |
| 704 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 705 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 706 | .channels(${CHANNEL_TILE}) |
| 707 | .qmax(192) |
| 708 | .Test(${", ".join(TEST_ARGS)}); |
| 709 | } |
| 710 | } |
| 711 | |
| 712 | $if CHANNEL_TILE > 1: |
| 713 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass) { |
| 714 | $if ISA_CHECK: |
| 715 | ${ISA_CHECK}; |
| 716 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 717 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 718 | MaxPoolMicrokernelTester() |
| 719 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 720 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 721 | .channels(channels) |
| 722 | .Test(${", ".join(TEST_ARGS)}); |
| 723 | } |
| 724 | } |
| 725 | } |
| 726 | |
| 727 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 728 | $if ISA_CHECK: |
| 729 | ${ISA_CHECK}; |
| 730 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 731 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 732 | MaxPoolMicrokernelTester() |
| 733 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 734 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 735 | .channels(channels) |
| 736 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 737 | .Test(${", ".join(TEST_ARGS)}); |
| 738 | } |
| 739 | } |
| 740 | } |
| 741 | |
| 742 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass_with_qmin) { |
| 743 | $if ISA_CHECK: |
| 744 | ${ISA_CHECK}; |
| 745 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 746 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 747 | MaxPoolMicrokernelTester() |
| 748 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 749 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 750 | .channels(channels) |
| 751 | .qmin(192) |
| 752 | .Test(${", ".join(TEST_ARGS)}); |
| 753 | } |
| 754 | } |
| 755 | } |
| 756 | |
| 757 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass_with_qmax) { |
| 758 | $if ISA_CHECK: |
| 759 | ${ISA_CHECK}; |
| 760 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 761 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 762 | MaxPoolMicrokernelTester() |
| 763 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 764 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 765 | .channels(channels) |
| 766 | .qmax(192) |
| 767 | .Test(${", ".join(TEST_ARGS)}); |
| 768 | } |
| 769 | } |
| 770 | } |
| 771 | |
| 772 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass) { |
| 773 | $if ISA_CHECK: |
| 774 | ${ISA_CHECK}; |
| 775 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 776 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 777 | MaxPoolMicrokernelTester() |
| 778 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 779 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 780 | .channels(channels) |
| 781 | .Test(${", ".join(TEST_ARGS)}); |
| 782 | } |
| 783 | } |
| 784 | } |
| 785 | |
| 786 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 787 | $if ISA_CHECK: |
| 788 | ${ISA_CHECK}; |
| 789 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 790 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 791 | MaxPoolMicrokernelTester() |
| 792 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 793 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 794 | .channels(channels) |
| 795 | .input_offset(${CHANNEL_TILE}) |
| 796 | .Test(${", ".join(TEST_ARGS)}); |
| 797 | } |
| 798 | } |
| 799 | } |
| 800 | |
| 801 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass_with_qmin) { |
| 802 | $if ISA_CHECK: |
| 803 | ${ISA_CHECK}; |
| 804 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 805 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 806 | MaxPoolMicrokernelTester() |
| 807 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 808 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 809 | .channels(channels) |
| 810 | .qmin(192) |
| 811 | .Test(${", ".join(TEST_ARGS)}); |
| 812 | } |
| 813 | } |
| 814 | } |
| 815 | |
| 816 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass_with_qmax) { |
| 817 | $if ISA_CHECK: |
| 818 | ${ISA_CHECK}; |
| 819 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 820 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 821 | MaxPoolMicrokernelTester() |
| 822 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 823 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 824 | .channels(channels) |
| 825 | .qmax(192) |
| 826 | .Test(${", ".join(TEST_ARGS)}); |
| 827 | } |
| 828 | } |
| 829 | } |
| 830 | |
| 831 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass) { |
| 832 | $if ISA_CHECK: |
| 833 | ${ISA_CHECK}; |
| 834 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 835 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 836 | MaxPoolMicrokernelTester() |
| 837 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 838 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 839 | .channels(channels) |
| 840 | .Test(${", ".join(TEST_ARGS)}); |
| 841 | } |
| 842 | } |
| 843 | } |
| 844 | |
| 845 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 846 | $if ISA_CHECK: |
| 847 | ${ISA_CHECK}; |
| 848 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 849 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 850 | MaxPoolMicrokernelTester() |
| 851 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 852 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 853 | .channels(channels) |
| 854 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 855 | .Test(${", ".join(TEST_ARGS)}); |
| 856 | } |
| 857 | } |
| 858 | } |
| 859 | |
| 860 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass_with_qmin) { |
| 861 | $if ISA_CHECK: |
| 862 | ${ISA_CHECK}; |
| 863 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 864 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 865 | MaxPoolMicrokernelTester() |
| 866 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 867 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 868 | .channels(channels) |
| 869 | .qmin(192) |
| 870 | .Test(${", ".join(TEST_ARGS)}); |
| 871 | } |
| 872 | } |
| 873 | } |
| 874 | |
| 875 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass_with_qmax) { |
| 876 | $if ISA_CHECK: |
| 877 | ${ISA_CHECK}; |
| 878 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 879 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 880 | MaxPoolMicrokernelTester() |
| 881 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 882 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 883 | .channels(channels) |
| 884 | .qmax(192) |
| 885 | .Test(${", ".join(TEST_ARGS)}); |
| 886 | } |
| 887 | } |
| 888 | } |
| 889 | |
| 890 | TEST(${TEST_NAME}, few_output_pixels) { |
| 891 | $if ISA_CHECK: |
| 892 | ${ISA_CHECK}; |
| 893 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 894 | for (size_t pooling_elements : std::vector<size_t>{{2, ${PRIMARY_TILE}, ${PRIMARY_TILE+INCREMENTAL_TILE-1}}}) { |
| 895 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 896 | MaxPoolMicrokernelTester() |
| 897 | .output_pixels(output_pixels) |
| 898 | .pooling_elements(pooling_elements) |
| 899 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 900 | .channels(channels) |
| 901 | .Test(${", ".join(TEST_ARGS)}); |
| 902 | } |
| 903 | } |
| 904 | } |
| 905 | } |
| 906 | |
| 907 | TEST(${TEST_NAME}, few_output_pixels_with_input_offset) { |
| 908 | $if ISA_CHECK: |
| 909 | ${ISA_CHECK}; |
| 910 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 911 | for (size_t pooling_elements : std::vector<size_t>{{2, ${PRIMARY_TILE}, ${PRIMARY_TILE+INCREMENTAL_TILE-1}}}) { |
| 912 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 913 | MaxPoolMicrokernelTester() |
| 914 | .output_pixels(output_pixels) |
| 915 | .pooling_elements(pooling_elements) |
| 916 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 917 | .channels(channels) |
| 918 | .input_offset(${next_prime(CHANNEL_TILE*5+1)}) |
| 919 | .Test(${", ".join(TEST_ARGS)}); |
| 920 | } |
| 921 | } |
| 922 | } |
| 923 | } |
| 924 | |
| 925 | TEST(${TEST_NAME}, few_output_pixels_with_qmin) { |
| 926 | $if ISA_CHECK: |
| 927 | ${ISA_CHECK}; |
| 928 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 929 | for (size_t pooling_elements : std::vector<size_t>{{2, ${PRIMARY_TILE}, ${PRIMARY_TILE+INCREMENTAL_TILE-1}}}) { |
| 930 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 931 | MaxPoolMicrokernelTester() |
| 932 | .output_pixels(output_pixels) |
| 933 | .pooling_elements(pooling_elements) |
| 934 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 935 | .channels(channels) |
| 936 | .qmin(192) |
| 937 | .Test(${", ".join(TEST_ARGS)}); |
| 938 | } |
| 939 | } |
| 940 | } |
| 941 | } |
| 942 | |
| 943 | TEST(${TEST_NAME}, few_output_pixels_with_qmax) { |
| 944 | $if ISA_CHECK: |
| 945 | ${ISA_CHECK}; |
| 946 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 947 | for (size_t pooling_elements : std::vector<size_t>{{2, ${PRIMARY_TILE}, ${PRIMARY_TILE+INCREMENTAL_TILE-1}}}) { |
| 948 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 949 | MaxPoolMicrokernelTester() |
| 950 | .output_pixels(output_pixels) |
| 951 | .pooling_elements(pooling_elements) |
| 952 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 953 | .channels(channels) |
| 954 | .qmax(192) |
| 955 | .Test(${", ".join(TEST_ARGS)}); |
| 956 | } |
| 957 | } |
| 958 | } |
| 959 | } |
| 960 | |
| 961 | TEST(${TEST_NAME}, few_output_pixels_with_output_stride) { |
| 962 | $if ISA_CHECK: |
| 963 | ${ISA_CHECK}; |
| 964 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 965 | for (size_t pooling_elements : std::vector<size_t>{{2, ${PRIMARY_TILE}, ${PRIMARY_TILE+INCREMENTAL_TILE-1}}}) { |
| 966 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 967 | MaxPoolMicrokernelTester() |
| 968 | .output_pixels(output_pixels) |
| 969 | .pooling_elements(pooling_elements) |
| 970 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 971 | .channels(channels) |
| 972 | .output_stride(${next_prime(CHANNEL_TILE*5+1)}) |
| 973 | .Test(${", ".join(TEST_ARGS)}); |
| 974 | } |
| 975 | } |
| 976 | } |
| 977 | } |
| 978 | |
| 979 | TEST(${TEST_NAME}, few_output_pixels_with_step) { |
| 980 | $if ISA_CHECK: |
| 981 | ${ISA_CHECK}; |
| 982 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 983 | for (size_t pooling_elements : std::vector<size_t>{{2, ${PRIMARY_TILE}, ${PRIMARY_TILE+INCREMENTAL_TILE-1}}}) { |
| 984 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 985 | for (size_t step = 2; step <= pooling_elements; step++) { |
| 986 | MaxPoolMicrokernelTester() |
| 987 | .output_pixels(output_pixels) |
| 988 | .pooling_elements(pooling_elements) |
| 989 | .pooling_tile(${PRIMARY_TILE}, ${INCREMENTAL_TILE}) |
| 990 | .step(step) |
| 991 | .channels(channels) |
| 992 | .output_stride(${next_prime(CHANNEL_TILE*5+1)}) |
| 993 | .Test(${", ".join(TEST_ARGS)}); |
| 994 | } |
| 995 | } |
| 996 | } |
| 997 | } |
| 998 | } |
| 999 | """ |
| 1000 | |
| 1001 | |
| 1002 | def generate_test_cases(ukernel, primary_tile, incremental_tile, channel_tile, |
| 1003 | isa): |
| 1004 | """Generates all tests cases for a MAXPOOL micro-kernel. |
| 1005 | |
| 1006 | Args: |
| 1007 | ukernel: C name of the micro-kernel function. |
| 1008 | primary_tile: Number of rows (pixels) processed per one iteration of the |
| 1009 | primary outer loop of the micro-kernel. |
| 1010 | incremental_tile: Number of rows (pixels) processed per one iteration of |
| 1011 | the incremental outer loop of the micro-kernel. |
| 1012 | channel_tile: Number of channels processed per one iteration of the inner |
| 1013 | loops of the micro-kernel. |
| 1014 | isa: instruction set required to run the micro-kernel. Generated unit test |
| 1015 | will skip execution if the host processor doesn't support this ISA. |
| 1016 | |
| 1017 | Returns: |
| 1018 | Code for the test case. |
| 1019 | """ |
| 1020 | _, test_name = ukernel.split("_", 1) |
| 1021 | _, datatype, ukernel_type, _ = ukernel.split("_", 3) |
| 1022 | test_args = [ukernel] |
| 1023 | if not isa or isa == "psimd": |
| 1024 | test_args.append("MaxPoolMicrokernelTester::Variant::Scalar") |
| 1025 | return xngen.preprocess(MAXPOOL_TEST_TEMPLATE, { |
| 1026 | "TEST_NAME": test_name.upper().replace("UKERNEL_", ""), |
| 1027 | "TEST_ARGS": test_args, |
| 1028 | "DATATYPE": datatype, |
| 1029 | "PRIMARY_TILE": primary_tile, |
| 1030 | "INCREMENTAL_TILE": incremental_tile, |
| 1031 | "CHANNEL_TILE": channel_tile, |
| 1032 | "ISA_CHECK": xnncommon.generate_isa_check_macro(isa), |
| 1033 | "next_prime": next_prime, |
| 1034 | }) |
| 1035 | |
| 1036 | |
| 1037 | def main(args): |
| 1038 | options = parser.parse_args(args) |
| 1039 | |
| 1040 | with codecs.open(options.spec, "r", encoding="utf-8") as spec_file: |
| 1041 | spec_yaml = yaml.safe_load(spec_file) |
| 1042 | if not isinstance(spec_yaml, list): |
| 1043 | raise ValueError("expected a list of micro-kernels in the spec") |
| 1044 | |
| 1045 | tests = """\ |
| 1046 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 1047 | // All rights reserved. |
| 1048 | // |
| 1049 | // Copyright 2019 Google LLC |
| 1050 | // |
| 1051 | // This source code is licensed under the BSD-style license found in the |
| 1052 | // LICENSE file in the root directory of this source tree. |
| 1053 | // |
| 1054 | // Auto-generated file. Do not edit! |
| 1055 | // Specification: {specification} |
| 1056 | // Generator: {generator} |
| 1057 | |
| 1058 | |
| 1059 | #include <gtest/gtest.h> |
| 1060 | |
| 1061 | #include <xnnpack/common.h> |
| 1062 | #include <xnnpack/isa-checks.h> |
| 1063 | |
| 1064 | #include <xnnpack/maxpool.h> |
| 1065 | #include "maxpool-microkernel-tester.h" |
| 1066 | """.format(specification=options.spec, generator=sys.argv[0]) |
| 1067 | |
| 1068 | for ukernel_spec in spec_yaml: |
| 1069 | name = ukernel_spec["name"] |
| 1070 | primary_tile, incremental_tile, channel_tile, arch, isa = \ |
| 1071 | split_ukernel_name(name) |
| 1072 | |
| 1073 | # specification can override architecture |
| 1074 | arch = ukernel_spec.get("arch", arch) |
| 1075 | |
| 1076 | test_case = generate_test_cases(name, primary_tile, incremental_tile, |
| 1077 | channel_tile, isa) |
| 1078 | tests += "\n\n" + xnncommon.postprocess_test_case(test_case, arch, isa) |
| 1079 | |
| 1080 | with codecs.open(options.output, "w", encoding="utf-8") as output_file: |
| 1081 | output_file.write(tests) |
| 1082 | |
| 1083 | |
| 1084 | if __name__ == "__main__": |
| 1085 | main(sys.argv[1:]) |