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( |
| 22 | description='ArgMaxPool microkernel test generator') |
| 23 | parser.add_argument("-s", "--spec", metavar="FILE", required=True, |
| 24 | help="Specification (YAML) file") |
| 25 | parser.add_argument("-o", "--output", metavar="FILE", required=True, |
| 26 | help='Output (C++ source) file') |
| 27 | parser.set_defaults(defines=list()) |
| 28 | |
| 29 | |
| 30 | def split_ukernel_name(name): |
| 31 | match = re.match(r"^xnn_(f16|f32)_argmaxpool_ukernel_((\d+)p)?(\d+)x__(.+)_c(\d+)$", name) |
| 32 | if match is None: |
| 33 | raise ValueError("Unexpected microkernel name: " + name) |
| 34 | |
| 35 | if match.group(2): |
| 36 | primary_tile = int(match.group(3)) |
| 37 | incremental_tile = int(match.group(4)) |
| 38 | else: |
| 39 | primary_tile = int(match.group(4)) |
| 40 | incremental_tile = 0 |
| 41 | |
| 42 | channel_tile = int(match.group(6)) |
| 43 | |
| 44 | arch, isa = xnncommon.parse_target_name(target_name=match.group(5)) |
| 45 | return primary_tile, incremental_tile, channel_tile, arch, isa |
| 46 | |
| 47 | |
| 48 | ARGMAXPOOL_TEST_TEMPLATE = """\ |
| 49 | $if INCREMENTAL_TILE == 0: |
| 50 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile) { |
| 51 | $if ISA_CHECK: |
| 52 | ${ISA_CHECK}; |
| 53 | ArgMaxPoolMicrokernelTester() |
| 54 | .pooling_elements(${PRIMARY_TILE}) |
| 55 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 56 | .channels(${CHANNEL_TILE}) |
| 57 | .Test(${", ".join(TEST_ARGS)}); |
| 58 | } |
| 59 | |
| 60 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 61 | $if ISA_CHECK: |
| 62 | ${ISA_CHECK}; |
| 63 | ArgMaxPoolMicrokernelTester() |
| 64 | .pooling_elements(${PRIMARY_TILE}) |
| 65 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 66 | .channels(${CHANNEL_TILE}) |
| 67 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 68 | .Test(${", ".join(TEST_ARGS)}); |
| 69 | } |
| 70 | |
| 71 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 72 | $if ISA_CHECK: |
| 73 | ${ISA_CHECK}; |
| 74 | ArgMaxPoolMicrokernelTester() |
| 75 | .pooling_elements(${PRIMARY_TILE}) |
| 76 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 77 | .channels(${CHANNEL_TILE}) |
| 78 | .qmin(192) |
| 79 | .Test(${", ".join(TEST_ARGS)}); |
| 80 | } |
| 81 | |
| 82 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 83 | $if ISA_CHECK: |
| 84 | ${ISA_CHECK}; |
| 85 | ArgMaxPoolMicrokernelTester() |
| 86 | .pooling_elements(${PRIMARY_TILE}) |
| 87 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 88 | .channels(${CHANNEL_TILE}) |
| 89 | .qmax(192) |
| 90 | .Test(${", ".join(TEST_ARGS)}); |
| 91 | } |
| 92 | |
| 93 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_subtile) { |
| 94 | $if ISA_CHECK: |
| 95 | ${ISA_CHECK}; |
| 96 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 97 | ArgMaxPoolMicrokernelTester() |
| 98 | .pooling_elements(pooling_elements) |
| 99 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 100 | .channels(${CHANNEL_TILE}) |
| 101 | .Test(${", ".join(TEST_ARGS)}); |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 106 | $if ISA_CHECK: |
| 107 | ${ISA_CHECK}; |
| 108 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 109 | ArgMaxPoolMicrokernelTester() |
| 110 | .pooling_elements(pooling_elements) |
| 111 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 112 | .channels(${CHANNEL_TILE}) |
| 113 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 114 | .Test(${", ".join(TEST_ARGS)}); |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | $if CHANNEL_TILE > 1: |
| 119 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile) { |
| 120 | $if ISA_CHECK: |
| 121 | ${ISA_CHECK}; |
| 122 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 123 | ArgMaxPoolMicrokernelTester() |
| 124 | .pooling_elements(${PRIMARY_TILE}) |
| 125 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 126 | .channels(channels) |
| 127 | .Test(${", ".join(TEST_ARGS)}); |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 132 | $if ISA_CHECK: |
| 133 | ${ISA_CHECK}; |
| 134 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 135 | ArgMaxPoolMicrokernelTester() |
| 136 | .pooling_elements(${PRIMARY_TILE}) |
| 137 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 138 | .channels(channels) |
| 139 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 140 | .Test(${", ".join(TEST_ARGS)}); |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 145 | $if ISA_CHECK: |
| 146 | ${ISA_CHECK}; |
| 147 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 148 | ArgMaxPoolMicrokernelTester() |
| 149 | .pooling_elements(${PRIMARY_TILE}) |
| 150 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 151 | .channels(channels) |
| 152 | .qmin(192) |
| 153 | .Test(${", ".join(TEST_ARGS)}); |
| 154 | } |
| 155 | } |
| 156 | |
| 157 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 158 | $if ISA_CHECK: |
| 159 | ${ISA_CHECK}; |
| 160 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 161 | ArgMaxPoolMicrokernelTester() |
| 162 | .pooling_elements(${PRIMARY_TILE}) |
| 163 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 164 | .channels(channels) |
| 165 | .qmax(192) |
| 166 | .Test(${", ".join(TEST_ARGS)}); |
| 167 | } |
| 168 | } |
| 169 | |
| 170 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_subtile) { |
| 171 | $if ISA_CHECK: |
| 172 | ${ISA_CHECK}; |
| 173 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 174 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 175 | ArgMaxPoolMicrokernelTester() |
| 176 | .pooling_elements(pooling_elements) |
| 177 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 178 | .channels(channels) |
| 179 | .Test(${", ".join(TEST_ARGS)}); |
| 180 | } |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 185 | $if ISA_CHECK: |
| 186 | ${ISA_CHECK}; |
| 187 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 188 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 189 | ArgMaxPoolMicrokernelTester() |
| 190 | .pooling_elements(pooling_elements) |
| 191 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 192 | .channels(channels) |
| 193 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 194 | .Test(${", ".join(TEST_ARGS)}); |
| 195 | } |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile) { |
| 200 | $if ISA_CHECK: |
| 201 | ${ISA_CHECK}; |
| 202 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 203 | ArgMaxPoolMicrokernelTester() |
| 204 | .pooling_elements(${PRIMARY_TILE}) |
| 205 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 206 | .channels(channels) |
| 207 | .Test(${", ".join(TEST_ARGS)}); |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 212 | $if ISA_CHECK: |
| 213 | ${ISA_CHECK}; |
| 214 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 215 | ArgMaxPoolMicrokernelTester() |
| 216 | .pooling_elements(${PRIMARY_TILE}) |
| 217 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 218 | .channels(channels) |
| 219 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 220 | .Test(${", ".join(TEST_ARGS)}); |
| 221 | } |
| 222 | } |
| 223 | |
| 224 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 225 | $if ISA_CHECK: |
| 226 | ${ISA_CHECK}; |
| 227 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 228 | ArgMaxPoolMicrokernelTester() |
| 229 | .pooling_elements(${PRIMARY_TILE}) |
| 230 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 231 | .channels(channels) |
| 232 | .qmin(192) |
| 233 | .Test(${", ".join(TEST_ARGS)}); |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 238 | $if ISA_CHECK: |
| 239 | ${ISA_CHECK}; |
| 240 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 241 | ArgMaxPoolMicrokernelTester() |
| 242 | .pooling_elements(${PRIMARY_TILE}) |
| 243 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 244 | .channels(channels) |
| 245 | .qmax(192) |
| 246 | .Test(${", ".join(TEST_ARGS)}); |
| 247 | } |
| 248 | } |
| 249 | |
| 250 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_subtile) { |
| 251 | $if ISA_CHECK: |
| 252 | ${ISA_CHECK}; |
| 253 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 254 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 255 | ArgMaxPoolMicrokernelTester() |
| 256 | .pooling_elements(pooling_elements) |
| 257 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 258 | .channels(channels) |
| 259 | .Test(${", ".join(TEST_ARGS)}); |
| 260 | } |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 265 | $if ISA_CHECK: |
| 266 | ${ISA_CHECK}; |
| 267 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 268 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 269 | ArgMaxPoolMicrokernelTester() |
| 270 | .pooling_elements(pooling_elements) |
| 271 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 272 | .channels(channels) |
| 273 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 274 | .Test(${", ".join(TEST_ARGS)}); |
| 275 | } |
| 276 | } |
| 277 | } |
| 278 | |
| 279 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile) { |
| 280 | $if ISA_CHECK: |
| 281 | ${ISA_CHECK}; |
| 282 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 283 | ArgMaxPoolMicrokernelTester() |
| 284 | .pooling_elements(${PRIMARY_TILE}) |
| 285 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 286 | .channels(channels) |
| 287 | .Test(${", ".join(TEST_ARGS)}); |
| 288 | } |
| 289 | } |
| 290 | |
| 291 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile_with_input_offset) { |
| 292 | $if ISA_CHECK: |
| 293 | ${ISA_CHECK}; |
| 294 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 295 | ArgMaxPoolMicrokernelTester() |
| 296 | .pooling_elements(${PRIMARY_TILE}) |
| 297 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 298 | .channels(channels) |
| 299 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 300 | .Test(${", ".join(TEST_ARGS)}); |
| 301 | } |
| 302 | } |
| 303 | |
| 304 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile_with_qmin) { |
| 305 | $if ISA_CHECK: |
| 306 | ${ISA_CHECK}; |
| 307 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 308 | ArgMaxPoolMicrokernelTester() |
| 309 | .pooling_elements(${PRIMARY_TILE}) |
| 310 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 311 | .channels(channels) |
| 312 | .qmin(192) |
| 313 | .Test(${", ".join(TEST_ARGS)}); |
| 314 | } |
| 315 | } |
| 316 | |
| 317 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_fulltile_with_qmax) { |
| 318 | $if ISA_CHECK: |
| 319 | ${ISA_CHECK}; |
| 320 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 321 | ArgMaxPoolMicrokernelTester() |
| 322 | .pooling_elements(${PRIMARY_TILE}) |
| 323 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 324 | .channels(channels) |
| 325 | .qmax(192) |
| 326 | .Test(${", ".join(TEST_ARGS)}); |
| 327 | } |
| 328 | } |
| 329 | |
| 330 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_subtile) { |
| 331 | $if ISA_CHECK: |
| 332 | ${ISA_CHECK}; |
| 333 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 334 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 335 | ArgMaxPoolMicrokernelTester() |
| 336 | .pooling_elements(pooling_elements) |
| 337 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 338 | .channels(channels) |
| 339 | .Test(${", ".join(TEST_ARGS)}); |
| 340 | } |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_unipass_subtile_with_input_offset) { |
| 345 | $if ISA_CHECK: |
| 346 | ${ISA_CHECK}; |
| 347 | for (size_t pooling_elements = 2; pooling_elements < ${PRIMARY_TILE}; pooling_elements++) { |
| 348 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 349 | ArgMaxPoolMicrokernelTester() |
| 350 | .pooling_elements(pooling_elements) |
| 351 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 352 | .channels(channels) |
| 353 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 354 | .Test(${", ".join(TEST_ARGS)}); |
| 355 | } |
| 356 | } |
| 357 | } |
| 358 | |
| 359 | $if INCREMENTAL_TILE != 0: |
| 360 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile) { |
| 361 | $if ISA_CHECK: |
| 362 | ${ISA_CHECK}; |
| 363 | ArgMaxPoolMicrokernelTester() |
| 364 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 365 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 366 | .channels(${CHANNEL_TILE}) |
| 367 | .Test(${", ".join(TEST_ARGS)}); |
| 368 | } |
| 369 | |
| 370 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 371 | $if ISA_CHECK: |
| 372 | ${ISA_CHECK}; |
| 373 | ArgMaxPoolMicrokernelTester() |
| 374 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 375 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 376 | .channels(${CHANNEL_TILE}) |
| 377 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 378 | .Test(${", ".join(TEST_ARGS)}); |
| 379 | } |
| 380 | |
| 381 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 382 | $if ISA_CHECK: |
| 383 | ${ISA_CHECK}; |
| 384 | ArgMaxPoolMicrokernelTester() |
| 385 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 386 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 387 | .channels(${CHANNEL_TILE}) |
| 388 | .qmin(192) |
| 389 | .Test(${", ".join(TEST_ARGS)}); |
| 390 | } |
| 391 | |
| 392 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 393 | $if ISA_CHECK: |
| 394 | ${ISA_CHECK}; |
| 395 | ArgMaxPoolMicrokernelTester() |
| 396 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 397 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 398 | .channels(${CHANNEL_TILE}) |
| 399 | .qmax(192) |
| 400 | .Test(${", ".join(TEST_ARGS)}); |
| 401 | } |
| 402 | |
| 403 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_subtile) { |
| 404 | $if ISA_CHECK: |
| 405 | ${ISA_CHECK}; |
| 406 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 407 | ArgMaxPoolMicrokernelTester() |
| 408 | .pooling_elements(pooling_elements) |
| 409 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 410 | .channels(${CHANNEL_TILE}) |
| 411 | .Test(${", ".join(TEST_ARGS)}); |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 416 | $if ISA_CHECK: |
| 417 | ${ISA_CHECK}; |
| 418 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 419 | ArgMaxPoolMicrokernelTester() |
| 420 | .pooling_elements(pooling_elements) |
| 421 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 422 | .channels(${CHANNEL_TILE}) |
| 423 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 424 | .Test(${", ".join(TEST_ARGS)}); |
| 425 | } |
| 426 | } |
| 427 | |
| 428 | $if CHANNEL_TILE > 1: |
| 429 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile) { |
| 430 | $if ISA_CHECK: |
| 431 | ${ISA_CHECK}; |
| 432 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 433 | ArgMaxPoolMicrokernelTester() |
| 434 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 435 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 436 | .channels(channels) |
| 437 | .Test(${", ".join(TEST_ARGS)}); |
| 438 | } |
| 439 | } |
| 440 | |
| 441 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 442 | $if ISA_CHECK: |
| 443 | ${ISA_CHECK}; |
| 444 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 445 | ArgMaxPoolMicrokernelTester() |
| 446 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 447 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 448 | .channels(channels) |
| 449 | .input_offset(${next_prime(CHANNEL_TILE*5)}) |
| 450 | .Test(${", ".join(TEST_ARGS)}); |
| 451 | } |
| 452 | } |
| 453 | |
| 454 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 455 | $if ISA_CHECK: |
| 456 | ${ISA_CHECK}; |
| 457 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 458 | ArgMaxPoolMicrokernelTester() |
| 459 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 460 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 461 | .channels(channels) |
| 462 | .qmin(192) |
| 463 | .Test(${", ".join(TEST_ARGS)}); |
| 464 | } |
| 465 | } |
| 466 | |
| 467 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 468 | $if ISA_CHECK: |
| 469 | ${ISA_CHECK}; |
| 470 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 471 | ArgMaxPoolMicrokernelTester() |
| 472 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 473 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 474 | .channels(channels) |
| 475 | .qmax(192) |
| 476 | .Test(${", ".join(TEST_ARGS)}); |
| 477 | } |
| 478 | } |
| 479 | |
| 480 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_subtile) { |
| 481 | $if ISA_CHECK: |
| 482 | ${ISA_CHECK}; |
| 483 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 484 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 485 | ArgMaxPoolMicrokernelTester() |
| 486 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 487 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 488 | .channels(channels) |
| 489 | .Test(${", ".join(TEST_ARGS)}); |
| 490 | } |
| 491 | } |
| 492 | } |
| 493 | |
| 494 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 495 | $if ISA_CHECK: |
| 496 | ${ISA_CHECK}; |
| 497 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 498 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 499 | ArgMaxPoolMicrokernelTester() |
| 500 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 501 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 502 | .channels(channels) |
| 503 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 504 | .Test(${", ".join(TEST_ARGS)}); |
| 505 | } |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile) { |
| 510 | $if ISA_CHECK: |
| 511 | ${ISA_CHECK}; |
| 512 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 513 | ArgMaxPoolMicrokernelTester() |
| 514 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 515 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 516 | .channels(channels) |
| 517 | .Test(${", ".join(TEST_ARGS)}); |
| 518 | } |
| 519 | } |
| 520 | |
| 521 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 522 | $if ISA_CHECK: |
| 523 | ${ISA_CHECK}; |
| 524 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 525 | ArgMaxPoolMicrokernelTester() |
| 526 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 527 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 528 | .channels(channels) |
| 529 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 530 | .Test(${", ".join(TEST_ARGS)}); |
| 531 | } |
| 532 | } |
| 533 | |
| 534 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 535 | $if ISA_CHECK: |
| 536 | ${ISA_CHECK}; |
| 537 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 538 | ArgMaxPoolMicrokernelTester() |
| 539 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 540 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 541 | .channels(channels) |
| 542 | .qmin(192) |
| 543 | .Test(${", ".join(TEST_ARGS)}); |
| 544 | } |
| 545 | } |
| 546 | |
| 547 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 548 | $if ISA_CHECK: |
| 549 | ${ISA_CHECK}; |
| 550 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 551 | ArgMaxPoolMicrokernelTester() |
| 552 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 553 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 554 | .channels(channels) |
| 555 | .qmax(192) |
| 556 | .Test(${", ".join(TEST_ARGS)}); |
| 557 | } |
| 558 | } |
| 559 | |
| 560 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_subtile) { |
| 561 | $if ISA_CHECK: |
| 562 | ${ISA_CHECK}; |
| 563 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 564 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 565 | ArgMaxPoolMicrokernelTester() |
| 566 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 567 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 568 | .channels(channels) |
| 569 | .Test(${", ".join(TEST_ARGS)}); |
| 570 | } |
| 571 | } |
| 572 | } |
| 573 | |
| 574 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 575 | $if ISA_CHECK: |
| 576 | ${ISA_CHECK}; |
| 577 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 578 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 579 | ArgMaxPoolMicrokernelTester() |
| 580 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 581 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 582 | .channels(channels) |
| 583 | .input_offset(${next_prime(CHANNEL_TILE)}) |
| 584 | .Test(${", ".join(TEST_ARGS)}); |
| 585 | } |
| 586 | } |
| 587 | } |
| 588 | |
| 589 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile) { |
| 590 | $if ISA_CHECK: |
| 591 | ${ISA_CHECK}; |
| 592 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 593 | ArgMaxPoolMicrokernelTester() |
| 594 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 595 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 596 | .channels(channels) |
| 597 | .Test(${", ".join(TEST_ARGS)}); |
| 598 | } |
| 599 | } |
| 600 | |
| 601 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile_with_input_offset) { |
| 602 | $if ISA_CHECK: |
| 603 | ${ISA_CHECK}; |
| 604 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 605 | ArgMaxPoolMicrokernelTester() |
| 606 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 607 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 608 | .channels(channels) |
| 609 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 610 | .Test(${", ".join(TEST_ARGS)}); |
| 611 | } |
| 612 | } |
| 613 | |
| 614 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile_with_qmin) { |
| 615 | $if ISA_CHECK: |
| 616 | ${ISA_CHECK}; |
| 617 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 618 | ArgMaxPoolMicrokernelTester() |
| 619 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 620 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 621 | .channels(channels) |
| 622 | .qmin(192) |
| 623 | .Test(${", ".join(TEST_ARGS)}); |
| 624 | } |
| 625 | } |
| 626 | |
| 627 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_fulltile_with_qmax) { |
| 628 | $if ISA_CHECK: |
| 629 | ${ISA_CHECK}; |
| 630 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 631 | ArgMaxPoolMicrokernelTester() |
| 632 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 633 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 634 | .channels(channels) |
| 635 | .qmax(192) |
| 636 | .Test(${", ".join(TEST_ARGS)}); |
| 637 | } |
| 638 | } |
| 639 | |
| 640 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_subtile) { |
| 641 | $if ISA_CHECK: |
| 642 | ${ISA_CHECK}; |
| 643 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 644 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 645 | ArgMaxPoolMicrokernelTester() |
| 646 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 647 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 648 | .channels(channels) |
| 649 | .Test(${", ".join(TEST_ARGS)}); |
| 650 | } |
| 651 | } |
| 652 | } |
| 653 | |
| 654 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_twopass_subtile_with_input_offset) { |
| 655 | $if ISA_CHECK: |
| 656 | ${ISA_CHECK}; |
| 657 | for (size_t pooling_elements = ${PRIMARY_TILE+1}; pooling_elements < ${PRIMARY_TILE+INCREMENTAL_TILE}; pooling_elements++) { |
| 658 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 659 | ArgMaxPoolMicrokernelTester() |
| 660 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 661 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 662 | .channels(channels) |
| 663 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 664 | .Test(${", ".join(TEST_ARGS)}); |
| 665 | } |
| 666 | } |
| 667 | } |
| 668 | |
| 669 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass) { |
| 670 | $if ISA_CHECK: |
| 671 | ${ISA_CHECK}; |
| 672 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 673 | ArgMaxPoolMicrokernelTester() |
| 674 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 675 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 676 | .channels(${CHANNEL_TILE}) |
| 677 | .Test(${", ".join(TEST_ARGS)}); |
| 678 | } |
| 679 | } |
| 680 | |
| 681 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 682 | $if ISA_CHECK: |
| 683 | ${ISA_CHECK}; |
| 684 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 685 | ArgMaxPoolMicrokernelTester() |
| 686 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 687 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 688 | .channels(${CHANNEL_TILE}) |
| 689 | .input_offset(${next_prime(CHANNEL_TILE+1)}) |
| 690 | .Test(${", ".join(TEST_ARGS)}); |
| 691 | } |
| 692 | } |
| 693 | |
| 694 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass_with_qmin) { |
| 695 | $if ISA_CHECK: |
| 696 | ${ISA_CHECK}; |
| 697 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 698 | ArgMaxPoolMicrokernelTester() |
| 699 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 700 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 701 | .channels(${CHANNEL_TILE}) |
| 702 | .qmin(192) |
| 703 | .Test(${", ".join(TEST_ARGS)}); |
| 704 | } |
| 705 | } |
| 706 | |
| 707 | TEST(${TEST_NAME}, channels_eq_${CHANNEL_TILE}_multipass_with_qmax) { |
| 708 | $if ISA_CHECK: |
| 709 | ${ISA_CHECK}; |
| 710 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 711 | ArgMaxPoolMicrokernelTester() |
| 712 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 713 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 714 | .channels(${CHANNEL_TILE}) |
| 715 | .qmax(192) |
| 716 | .Test(${", ".join(TEST_ARGS)}); |
| 717 | } |
| 718 | } |
| 719 | |
| 720 | $if CHANNEL_TILE > 1: |
| 721 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass) { |
| 722 | $if ISA_CHECK: |
| 723 | ${ISA_CHECK}; |
| 724 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 725 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 726 | ArgMaxPoolMicrokernelTester() |
| 727 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 728 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 729 | .channels(channels) |
| 730 | .Test(${", ".join(TEST_ARGS)}); |
| 731 | } |
| 732 | } |
| 733 | } |
| 734 | |
| 735 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 736 | $if ISA_CHECK: |
| 737 | ${ISA_CHECK}; |
| 738 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 739 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 740 | ArgMaxPoolMicrokernelTester() |
| 741 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 742 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 743 | .channels(channels) |
| 744 | .input_offset(${next_prime(CHANNEL_TILE*8)}) |
| 745 | .Test(${", ".join(TEST_ARGS)}); |
| 746 | } |
| 747 | } |
| 748 | } |
| 749 | |
| 750 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass_with_qmin) { |
| 751 | $if ISA_CHECK: |
| 752 | ${ISA_CHECK}; |
| 753 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 754 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 755 | ArgMaxPoolMicrokernelTester() |
| 756 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 757 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 758 | .channels(channels) |
| 759 | .qmin(192) |
| 760 | .Test(${", ".join(TEST_ARGS)}); |
| 761 | } |
| 762 | } |
| 763 | } |
| 764 | |
| 765 | TEST(${TEST_NAME}, channels_div_${CHANNEL_TILE}_multipass_with_qmax) { |
| 766 | $if ISA_CHECK: |
| 767 | ${ISA_CHECK}; |
| 768 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 769 | for (size_t channels = ${CHANNEL_TILE*2}; channels < ${CHANNEL_TILE*8}; channels += ${CHANNEL_TILE}) { |
| 770 | ArgMaxPoolMicrokernelTester() |
| 771 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 772 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 773 | .channels(channels) |
| 774 | .qmax(192) |
| 775 | .Test(${", ".join(TEST_ARGS)}); |
| 776 | } |
| 777 | } |
| 778 | } |
| 779 | |
| 780 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass) { |
| 781 | $if ISA_CHECK: |
| 782 | ${ISA_CHECK}; |
| 783 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 784 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 785 | ArgMaxPoolMicrokernelTester() |
| 786 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 787 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 788 | .channels(channels) |
| 789 | .Test(${", ".join(TEST_ARGS)}); |
| 790 | } |
| 791 | } |
| 792 | } |
| 793 | |
| 794 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 795 | $if ISA_CHECK: |
| 796 | ${ISA_CHECK}; |
| 797 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 798 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 799 | ArgMaxPoolMicrokernelTester() |
| 800 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 801 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 802 | .channels(channels) |
| 803 | .input_offset(${CHANNEL_TILE}) |
| 804 | .Test(${", ".join(TEST_ARGS)}); |
| 805 | } |
| 806 | } |
| 807 | } |
| 808 | |
| 809 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass_with_qmin) { |
| 810 | $if ISA_CHECK: |
| 811 | ${ISA_CHECK}; |
| 812 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 813 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 814 | ArgMaxPoolMicrokernelTester() |
| 815 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 816 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 817 | .channels(channels) |
| 818 | .qmin(192) |
| 819 | .Test(${", ".join(TEST_ARGS)}); |
| 820 | } |
| 821 | } |
| 822 | } |
| 823 | |
| 824 | TEST(${TEST_NAME}, channels_lt_${CHANNEL_TILE}_multipass_with_qmax) { |
| 825 | $if ISA_CHECK: |
| 826 | ${ISA_CHECK}; |
| 827 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 828 | for (size_t channels = 1; channels < ${CHANNEL_TILE}; channels++) { |
| 829 | ArgMaxPoolMicrokernelTester() |
| 830 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 831 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 832 | .channels(channels) |
| 833 | .qmax(192) |
| 834 | .Test(${", ".join(TEST_ARGS)}); |
| 835 | } |
| 836 | } |
| 837 | } |
| 838 | |
| 839 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass) { |
| 840 | $if ISA_CHECK: |
| 841 | ${ISA_CHECK}; |
| 842 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 843 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 844 | ArgMaxPoolMicrokernelTester() |
| 845 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 846 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 847 | .channels(channels) |
| 848 | .Test(${", ".join(TEST_ARGS)}); |
| 849 | } |
| 850 | } |
| 851 | } |
| 852 | |
| 853 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass_with_input_offset) { |
| 854 | $if ISA_CHECK: |
| 855 | ${ISA_CHECK}; |
| 856 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 857 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 858 | ArgMaxPoolMicrokernelTester() |
| 859 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 860 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 861 | .channels(channels) |
| 862 | .input_offset(${next_prime(CHANNEL_TILE*2)}) |
| 863 | .Test(${", ".join(TEST_ARGS)}); |
| 864 | } |
| 865 | } |
| 866 | } |
| 867 | |
| 868 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass_with_qmin) { |
| 869 | $if ISA_CHECK: |
| 870 | ${ISA_CHECK}; |
| 871 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 872 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 873 | ArgMaxPoolMicrokernelTester() |
| 874 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 875 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 876 | .channels(channels) |
| 877 | .qmin(192) |
| 878 | .Test(${", ".join(TEST_ARGS)}); |
| 879 | } |
| 880 | } |
| 881 | } |
| 882 | |
| 883 | TEST(${TEST_NAME}, channels_gt_${CHANNEL_TILE}_multipass_with_qmax) { |
| 884 | $if ISA_CHECK: |
| 885 | ${ISA_CHECK}; |
| 886 | for (size_t pooling_elements = ${PRIMARY_TILE+INCREMENTAL_TILE+1}; pooling_elements <= ${PRIMARY_TILE+INCREMENTAL_TILE*3}; pooling_elements += 3) { |
| 887 | for (size_t channels = ${CHANNEL_TILE+1}; channels < ${10 if CHANNEL_TILE == 1 else CHANNEL_TILE*2}; channels++) { |
| 888 | ArgMaxPoolMicrokernelTester() |
| 889 | .pooling_elements(${PRIMARY_TILE+INCREMENTAL_TILE}) |
| 890 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 891 | .channels(channels) |
| 892 | .qmax(192) |
| 893 | .Test(${", ".join(TEST_ARGS)}); |
| 894 | } |
| 895 | } |
| 896 | } |
| 897 | |
| 898 | $if INCREMENTAL_TILE == 0: |
| 899 | $MIN_POOLING, MAX_POOLING = 2, PRIMARY_TILE |
| 900 | $else: |
| 901 | $MIN_POOLING, MAX_POOLING = PRIMARY_TILE + 1, PRIMARY_TILE + INCREMENTAL_TILE |
| 902 | |
| 903 | TEST(${TEST_NAME}, few_output_pixels) { |
| 904 | $if ISA_CHECK: |
| 905 | ${ISA_CHECK}; |
| 906 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 907 | for (size_t pooling_elements = ${MIN_POOLING}; pooling_elements <= ${MAX_POOLING}; pooling_elements++) { |
| 908 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 909 | ArgMaxPoolMicrokernelTester() |
| 910 | .output_pixels(output_pixels) |
| 911 | .pooling_elements(pooling_elements) |
| 912 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 913 | .channels(channels) |
| 914 | .Test(${", ".join(TEST_ARGS)}); |
| 915 | } |
| 916 | } |
| 917 | } |
| 918 | } |
| 919 | |
| 920 | TEST(${TEST_NAME}, few_output_pixels_with_input_offset) { |
| 921 | $if ISA_CHECK: |
| 922 | ${ISA_CHECK}; |
| 923 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 924 | for (size_t pooling_elements = ${MIN_POOLING}; pooling_elements <= ${MAX_POOLING}; pooling_elements++) { |
| 925 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 926 | ArgMaxPoolMicrokernelTester() |
| 927 | .output_pixels(output_pixels) |
| 928 | .pooling_elements(pooling_elements) |
| 929 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 930 | .channels(channels) |
| 931 | .input_offset(${next_prime(CHANNEL_TILE*5+1)}) |
| 932 | .Test(${", ".join(TEST_ARGS)}); |
| 933 | } |
| 934 | } |
| 935 | } |
| 936 | } |
| 937 | |
| 938 | TEST(${TEST_NAME}, few_output_pixels_with_qmin) { |
| 939 | $if ISA_CHECK: |
| 940 | ${ISA_CHECK}; |
| 941 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 942 | for (size_t pooling_elements = ${MIN_POOLING}; pooling_elements <= ${MAX_POOLING}; pooling_elements++) { |
| 943 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 944 | ArgMaxPoolMicrokernelTester() |
| 945 | .output_pixels(output_pixels) |
| 946 | .pooling_elements(pooling_elements) |
| 947 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 948 | .channels(channels) |
| 949 | .qmin(192) |
| 950 | .Test(${", ".join(TEST_ARGS)}); |
| 951 | } |
| 952 | } |
| 953 | } |
| 954 | } |
| 955 | |
| 956 | TEST(${TEST_NAME}, few_output_pixels_with_qmax) { |
| 957 | $if ISA_CHECK: |
| 958 | ${ISA_CHECK}; |
| 959 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 960 | for (size_t pooling_elements = ${MIN_POOLING}; pooling_elements <= ${MAX_POOLING}; pooling_elements++) { |
| 961 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 962 | ArgMaxPoolMicrokernelTester() |
| 963 | .output_pixels(output_pixels) |
| 964 | .pooling_elements(pooling_elements) |
| 965 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 966 | .channels(channels) |
| 967 | .qmax(192) |
| 968 | .Test(${", ".join(TEST_ARGS)}); |
| 969 | } |
| 970 | } |
| 971 | } |
| 972 | } |
| 973 | |
| 974 | TEST(${TEST_NAME}, few_output_pixels_with_output_stride) { |
| 975 | $if ISA_CHECK: |
| 976 | ${ISA_CHECK}; |
| 977 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 978 | for (size_t pooling_elements = ${MIN_POOLING}; pooling_elements <= ${MAX_POOLING}; pooling_elements++) { |
| 979 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 980 | ArgMaxPoolMicrokernelTester() |
| 981 | .output_pixels(output_pixels) |
| 982 | .pooling_elements(pooling_elements) |
| 983 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 984 | .channels(channels) |
| 985 | .output_stride(${next_prime(CHANNEL_TILE*5+1)}) |
| 986 | .Test(${", ".join(TEST_ARGS)}); |
| 987 | } |
| 988 | } |
| 989 | } |
| 990 | } |
| 991 | |
| 992 | TEST(${TEST_NAME}, few_output_pixels_with_step) { |
| 993 | $if ISA_CHECK: |
| 994 | ${ISA_CHECK}; |
| 995 | for (size_t output_pixels = 2; output_pixels <= 5; output_pixels++) { |
| 996 | for (size_t pooling_elements = ${MIN_POOLING}; pooling_elements <= ${MAX_POOLING}; pooling_elements++) { |
| 997 | for (size_t channels = 1; channels <= ${CHANNEL_TILE*5}; channels += ${max(1, CHANNEL_TILE-1)}) { |
| 998 | for (size_t step = 2; step <= pooling_elements; step++) { |
| 999 | ArgMaxPoolMicrokernelTester() |
| 1000 | .output_pixels(output_pixels) |
| 1001 | .pooling_elements(pooling_elements) |
| 1002 | .pooling_tile(${", ".join(map(str, filter(bool, [PRIMARY_TILE, INCREMENTAL_TILE])))}) |
| 1003 | .step(step) |
| 1004 | .channels(channels) |
| 1005 | .output_stride(${next_prime(CHANNEL_TILE*5+1)}) |
| 1006 | .Test(${", ".join(TEST_ARGS)}); |
| 1007 | } |
| 1008 | } |
| 1009 | } |
| 1010 | } |
| 1011 | } |
| 1012 | """ |
| 1013 | |
| 1014 | |
| 1015 | def generate_test_cases(ukernel, primary_tile, incremental_tile, channel_tile, |
| 1016 | isa): |
| 1017 | """Generates all tests cases for a ARGMAXPOOL micro-kernel. |
| 1018 | |
| 1019 | Args: |
| 1020 | ukernel: C name of the micro-kernel function. |
| 1021 | primary_tile: Number of rows (pixels) processed per one iteration of the |
| 1022 | primary outer loop of the micro-kernel. |
| 1023 | incremental_tile: Number of rows (pixels) processed per one iteration of |
| 1024 | the incremental outer loop of the micro-kernel. |
| 1025 | channel_tile: Number of channels processed per one iteration of the inner |
| 1026 | loops of the micro-kernel. |
| 1027 | isa: instruction set required to run the micro-kernel. Generated unit test |
| 1028 | will skip execution if the host processor doesn't support this ISA. |
| 1029 | |
| 1030 | Returns: |
| 1031 | Code for the test case. |
| 1032 | """ |
| 1033 | _, test_name = ukernel.split("_", 1) |
| 1034 | _, datatype, ukernel_type, _ = ukernel.split("_", 3) |
| 1035 | test_args = [ukernel] |
| 1036 | if not isa or isa == "psimd": |
| 1037 | test_args.append("ArgMaxPoolMicrokernelTester::Variant::Scalar") |
| 1038 | return xngen.preprocess(ARGMAXPOOL_TEST_TEMPLATE, { |
| 1039 | "TEST_NAME": test_name.upper().replace("UKERNEL_", ""), |
| 1040 | "TEST_ARGS": test_args, |
| 1041 | "DATATYPE": datatype, |
| 1042 | "PRIMARY_TILE": primary_tile, |
| 1043 | "INCREMENTAL_TILE": incremental_tile, |
| 1044 | "CHANNEL_TILE": channel_tile, |
| 1045 | "ISA_CHECK": xnncommon.generate_isa_check_macro(isa), |
| 1046 | "next_prime": next_prime, |
| 1047 | }) |
| 1048 | |
| 1049 | |
| 1050 | def main(args): |
| 1051 | options = parser.parse_args(args) |
| 1052 | |
| 1053 | with codecs.open(options.spec, "r", encoding="utf-8") as spec_file: |
| 1054 | spec_yaml = yaml.safe_load(spec_file) |
| 1055 | if not isinstance(spec_yaml, list): |
| 1056 | raise ValueError("expected a list of micro-kernels in the spec") |
| 1057 | |
| 1058 | tests = """\ |
| 1059 | // Copyright 2019 Google LLC |
| 1060 | // |
| 1061 | // This source code is licensed under the BSD-style license found in the |
| 1062 | // LICENSE file in the root directory of this source tree. |
| 1063 | // |
| 1064 | // Auto-generated file. Do not edit! |
| 1065 | // Specification: {specification} |
| 1066 | // Generator: {generator} |
| 1067 | |
| 1068 | |
| 1069 | #include <gtest/gtest.h> |
| 1070 | |
| 1071 | #include <xnnpack/common.h> |
| 1072 | #include <xnnpack/isa-checks.h> |
| 1073 | |
| 1074 | #include <xnnpack/argmaxpool.h> |
| 1075 | #include "argmaxpool-microkernel-tester.h" |
| 1076 | """.format(specification=options.spec, generator=sys.argv[0]) |
| 1077 | |
| 1078 | for ukernel_spec in spec_yaml: |
| 1079 | name = ukernel_spec["name"] |
| 1080 | primary_tile, incremental_tile, channel_tile, arch, isa = \ |
| 1081 | split_ukernel_name(name) |
| 1082 | |
| 1083 | # specification can override architecture |
| 1084 | arch = ukernel_spec.get("arch", arch) |
| 1085 | |
| 1086 | test_case = generate_test_cases(name, primary_tile, incremental_tile, |
| 1087 | channel_tile, isa) |
| 1088 | tests += "\n\n" + xnncommon.postprocess_test_case(test_case, arch, isa) |
| 1089 | |
| 1090 | with codecs.open(options.output, "w", encoding="utf-8") as output_file: |
| 1091 | output_file.write(tests) |
| 1092 | |
| 1093 | |
| 1094 | if __name__ == "__main__": |
| 1095 | main(sys.argv[1:]) |