Marat Dukhan | 173661d | 2021-07-26 23:47:08 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | // |
| 9 | // Auto-generated file. Do not edit! |
| 10 | // Specification: test/qu8-gemm-minmax-rndnu.yaml |
| 11 | // Generator: tools/generate-gemm-test.py |
| 12 | |
| 13 | |
| 14 | #include <gtest/gtest.h> |
| 15 | |
Frank Barchard | 447aa7b | 2021-12-28 14:11:40 -0800 | [diff] [blame] | 16 | #include <xnnpack/allocator.h> |
Marat Dukhan | 173661d | 2021-07-26 23:47:08 -0700 | [diff] [blame] | 17 | #include <xnnpack/common.h> |
| 18 | #include <xnnpack/isa-checks.h> |
| 19 | |
| 20 | #include <xnnpack/gemm.h> |
| 21 | #include <xnnpack/igemm.h> |
| 22 | #include <xnnpack/ppmm.h> |
| 23 | #include "gemm-microkernel-tester.h" |
| 24 | |
| 25 | |
Frank Barchard | 9e4d2aa | 2022-02-02 00:31:21 -0800 | [diff] [blame] | 26 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 27 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) { |
| 28 | TEST_REQUIRES_ARM_NEON; |
| 29 | GemmMicrokernelTester() |
| 30 | .mr(4) |
| 31 | .nr(8) |
| 32 | .kr(1) |
| 33 | .sr(1) |
| 34 | .m(4) |
| 35 | .n(8) |
| 36 | .k(8) |
| 37 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 38 | } |
| 39 | |
| 40 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, strided_cn) { |
| 41 | TEST_REQUIRES_ARM_NEON; |
| 42 | GemmMicrokernelTester() |
| 43 | .mr(4) |
| 44 | .nr(8) |
| 45 | .kr(1) |
| 46 | .sr(1) |
| 47 | .m(4) |
| 48 | .n(8) |
| 49 | .k(8) |
| 50 | .cn_stride(11) |
| 51 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 52 | } |
| 53 | |
| 54 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_strided_a) { |
| 55 | TEST_REQUIRES_ARM_NEON; |
| 56 | GemmMicrokernelTester() |
| 57 | .mr(4) |
| 58 | .nr(8) |
| 59 | .kr(1) |
| 60 | .sr(1) |
| 61 | .m(4) |
| 62 | .n(8) |
| 63 | .k(8) |
| 64 | .a_stride(11) |
| 65 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 66 | } |
| 67 | |
| 68 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_subtile) { |
| 69 | TEST_REQUIRES_ARM_NEON; |
| 70 | for (uint32_t n = 1; n <= 8; n++) { |
| 71 | for (uint32_t m = 1; m <= 4; m++) { |
| 72 | GemmMicrokernelTester() |
| 73 | .mr(4) |
| 74 | .nr(8) |
| 75 | .kr(1) |
| 76 | .sr(1) |
| 77 | .m(m) |
| 78 | .n(n) |
| 79 | .k(8) |
| 80 | .iterations(1) |
| 81 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 82 | } |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_subtile_m) { |
| 87 | TEST_REQUIRES_ARM_NEON; |
| 88 | for (uint32_t m = 1; m <= 4; m++) { |
| 89 | GemmMicrokernelTester() |
| 90 | .mr(4) |
| 91 | .nr(8) |
| 92 | .kr(1) |
| 93 | .sr(1) |
| 94 | .m(m) |
| 95 | .n(8) |
| 96 | .k(8) |
| 97 | .iterations(1) |
| 98 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_subtile_n) { |
| 103 | TEST_REQUIRES_ARM_NEON; |
| 104 | for (uint32_t n = 1; n <= 8; n++) { |
| 105 | GemmMicrokernelTester() |
| 106 | .mr(4) |
| 107 | .nr(8) |
| 108 | .kr(1) |
| 109 | .sr(1) |
| 110 | .m(4) |
| 111 | .n(n) |
| 112 | .k(8) |
| 113 | .iterations(1) |
| 114 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_lt_8) { |
| 119 | TEST_REQUIRES_ARM_NEON; |
| 120 | for (size_t k = 1; k < 8; k++) { |
| 121 | GemmMicrokernelTester() |
| 122 | .mr(4) |
| 123 | .nr(8) |
| 124 | .kr(1) |
| 125 | .sr(1) |
| 126 | .m(4) |
| 127 | .n(8) |
| 128 | .k(k) |
| 129 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_lt_8_strided_a) { |
| 134 | TEST_REQUIRES_ARM_NEON; |
| 135 | for (size_t k = 1; k < 8; k++) { |
| 136 | GemmMicrokernelTester() |
| 137 | .mr(4) |
| 138 | .nr(8) |
| 139 | .kr(1) |
| 140 | .sr(1) |
| 141 | .m(4) |
| 142 | .n(8) |
| 143 | .k(k) |
| 144 | .a_stride(11) |
| 145 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_lt_8_subtile) { |
| 150 | TEST_REQUIRES_ARM_NEON; |
| 151 | for (size_t k = 1; k < 8; k++) { |
| 152 | for (uint32_t n = 1; n <= 8; n++) { |
| 153 | for (uint32_t m = 1; m <= 4; m++) { |
| 154 | GemmMicrokernelTester() |
| 155 | .mr(4) |
| 156 | .nr(8) |
| 157 | .kr(1) |
| 158 | .sr(1) |
| 159 | .m(m) |
| 160 | .n(n) |
| 161 | .k(k) |
| 162 | .iterations(1) |
| 163 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_gt_8) { |
| 170 | TEST_REQUIRES_ARM_NEON; |
| 171 | for (size_t k = 9; k < 16; k++) { |
| 172 | GemmMicrokernelTester() |
| 173 | .mr(4) |
| 174 | .nr(8) |
| 175 | .kr(1) |
| 176 | .sr(1) |
| 177 | .m(4) |
| 178 | .n(8) |
| 179 | .k(k) |
| 180 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_gt_8_strided_a) { |
| 185 | TEST_REQUIRES_ARM_NEON; |
| 186 | for (size_t k = 9; k < 16; k++) { |
| 187 | GemmMicrokernelTester() |
| 188 | .mr(4) |
| 189 | .nr(8) |
| 190 | .kr(1) |
| 191 | .sr(1) |
| 192 | .m(4) |
| 193 | .n(8) |
| 194 | .k(k) |
| 195 | .a_stride(19) |
| 196 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_gt_8_subtile) { |
| 201 | TEST_REQUIRES_ARM_NEON; |
| 202 | for (size_t k = 9; k < 16; k++) { |
| 203 | for (uint32_t n = 1; n <= 8; n++) { |
| 204 | for (uint32_t m = 1; m <= 4; m++) { |
| 205 | GemmMicrokernelTester() |
| 206 | .mr(4) |
| 207 | .nr(8) |
| 208 | .kr(1) |
| 209 | .sr(1) |
| 210 | .m(m) |
| 211 | .n(n) |
| 212 | .k(k) |
| 213 | .iterations(1) |
| 214 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_div_8) { |
| 221 | TEST_REQUIRES_ARM_NEON; |
| 222 | for (size_t k = 16; k <= 80; k += 8) { |
| 223 | GemmMicrokernelTester() |
| 224 | .mr(4) |
| 225 | .nr(8) |
| 226 | .kr(1) |
| 227 | .sr(1) |
| 228 | .m(4) |
| 229 | .n(8) |
| 230 | .k(k) |
| 231 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 232 | } |
| 233 | } |
| 234 | |
| 235 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_div_8_strided_a) { |
| 236 | TEST_REQUIRES_ARM_NEON; |
| 237 | for (size_t k = 16; k <= 80; k += 8) { |
| 238 | GemmMicrokernelTester() |
| 239 | .mr(4) |
| 240 | .nr(8) |
| 241 | .kr(1) |
| 242 | .sr(1) |
| 243 | .m(4) |
| 244 | .n(8) |
| 245 | .k(k) |
| 246 | .a_stride(83) |
| 247 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_div_8_subtile) { |
| 252 | TEST_REQUIRES_ARM_NEON; |
| 253 | for (size_t k = 16; k <= 80; k += 8) { |
| 254 | for (uint32_t n = 1; n <= 8; n++) { |
| 255 | for (uint32_t m = 1; m <= 4; m++) { |
| 256 | GemmMicrokernelTester() |
| 257 | .mr(4) |
| 258 | .nr(8) |
| 259 | .kr(1) |
| 260 | .sr(1) |
| 261 | .m(m) |
| 262 | .n(n) |
| 263 | .k(k) |
| 264 | .iterations(1) |
| 265 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 266 | } |
| 267 | } |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_8) { |
| 272 | TEST_REQUIRES_ARM_NEON; |
| 273 | for (uint32_t n = 9; n < 16; n++) { |
| 274 | for (size_t k = 1; k <= 40; k += 9) { |
| 275 | GemmMicrokernelTester() |
| 276 | .mr(4) |
| 277 | .nr(8) |
| 278 | .kr(1) |
| 279 | .sr(1) |
| 280 | .m(4) |
| 281 | .n(n) |
| 282 | .k(k) |
| 283 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 284 | } |
| 285 | } |
| 286 | } |
| 287 | |
| 288 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_8_strided_cn) { |
| 289 | TEST_REQUIRES_ARM_NEON; |
| 290 | for (uint32_t n = 9; n < 16; n++) { |
| 291 | for (size_t k = 1; k <= 40; k += 9) { |
| 292 | GemmMicrokernelTester() |
| 293 | .mr(4) |
| 294 | .nr(8) |
| 295 | .kr(1) |
| 296 | .sr(1) |
| 297 | .m(4) |
| 298 | .n(n) |
| 299 | .k(k) |
| 300 | .cn_stride(11) |
| 301 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 302 | } |
| 303 | } |
| 304 | } |
| 305 | |
| 306 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_8_strided_a) { |
| 307 | TEST_REQUIRES_ARM_NEON; |
| 308 | for (uint32_t n = 9; n < 16; n++) { |
| 309 | for (size_t k = 1; k <= 40; k += 9) { |
| 310 | GemmMicrokernelTester() |
| 311 | .mr(4) |
| 312 | .nr(8) |
| 313 | .kr(1) |
| 314 | .sr(1) |
| 315 | .m(4) |
| 316 | .n(n) |
| 317 | .k(k) |
| 318 | .a_stride(43) |
| 319 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 320 | } |
| 321 | } |
| 322 | } |
| 323 | |
| 324 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_8_subtile) { |
| 325 | TEST_REQUIRES_ARM_NEON; |
| 326 | for (uint32_t n = 9; n < 16; n++) { |
| 327 | for (size_t k = 1; k <= 40; k += 9) { |
| 328 | for (uint32_t m = 1; m <= 4; m++) { |
| 329 | GemmMicrokernelTester() |
| 330 | .mr(4) |
| 331 | .nr(8) |
| 332 | .kr(1) |
| 333 | .sr(1) |
| 334 | .m(m) |
| 335 | .n(n) |
| 336 | .k(k) |
| 337 | .iterations(1) |
| 338 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_8) { |
| 345 | TEST_REQUIRES_ARM_NEON; |
| 346 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 347 | for (size_t k = 1; k <= 40; k += 9) { |
| 348 | GemmMicrokernelTester() |
| 349 | .mr(4) |
| 350 | .nr(8) |
| 351 | .kr(1) |
| 352 | .sr(1) |
| 353 | .m(4) |
| 354 | .n(n) |
| 355 | .k(k) |
| 356 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 357 | } |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_8_strided_cn) { |
| 362 | TEST_REQUIRES_ARM_NEON; |
| 363 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 364 | for (size_t k = 1; k <= 40; k += 9) { |
| 365 | GemmMicrokernelTester() |
| 366 | .mr(4) |
| 367 | .nr(8) |
| 368 | .kr(1) |
| 369 | .sr(1) |
| 370 | .m(4) |
| 371 | .n(n) |
| 372 | .k(k) |
| 373 | .cn_stride(11) |
| 374 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 375 | } |
| 376 | } |
| 377 | } |
| 378 | |
| 379 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_8_strided_a) { |
| 380 | TEST_REQUIRES_ARM_NEON; |
| 381 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 382 | for (size_t k = 1; k <= 40; k += 9) { |
| 383 | GemmMicrokernelTester() |
| 384 | .mr(4) |
| 385 | .nr(8) |
| 386 | .kr(1) |
| 387 | .sr(1) |
| 388 | .m(4) |
| 389 | .n(n) |
| 390 | .k(k) |
| 391 | .a_stride(43) |
| 392 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 393 | } |
| 394 | } |
| 395 | } |
| 396 | |
| 397 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_8_subtile) { |
| 398 | TEST_REQUIRES_ARM_NEON; |
| 399 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 400 | for (size_t k = 1; k <= 40; k += 9) { |
| 401 | for (uint32_t m = 1; m <= 4; m++) { |
| 402 | GemmMicrokernelTester() |
| 403 | .mr(4) |
| 404 | .nr(8) |
| 405 | .kr(1) |
| 406 | .sr(1) |
| 407 | .m(m) |
| 408 | .n(n) |
| 409 | .k(k) |
| 410 | .iterations(1) |
| 411 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 412 | } |
| 413 | } |
| 414 | } |
| 415 | } |
| 416 | |
| 417 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, strided_cm_subtile) { |
| 418 | TEST_REQUIRES_ARM_NEON; |
| 419 | for (size_t k = 1; k <= 40; k += 9) { |
| 420 | for (uint32_t n = 1; n <= 8; n++) { |
| 421 | for (uint32_t m = 1; m <= 4; m++) { |
| 422 | GemmMicrokernelTester() |
| 423 | .mr(4) |
| 424 | .nr(8) |
| 425 | .kr(1) |
| 426 | .sr(1) |
| 427 | .m(m) |
| 428 | .n(n) |
| 429 | .k(k) |
| 430 | .cm_stride(11) |
| 431 | .iterations(1) |
| 432 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 433 | } |
| 434 | } |
| 435 | } |
| 436 | } |
| 437 | |
| 438 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, qmin) { |
| 439 | TEST_REQUIRES_ARM_NEON; |
| 440 | GemmMicrokernelTester() |
| 441 | .mr(4) |
| 442 | .nr(8) |
| 443 | .kr(1) |
| 444 | .sr(1) |
| 445 | .m(4) |
| 446 | .n(8) |
| 447 | .k(8) |
| 448 | .qmin(128) |
| 449 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 450 | } |
| 451 | |
| 452 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, qmax) { |
| 453 | TEST_REQUIRES_ARM_NEON; |
| 454 | GemmMicrokernelTester() |
| 455 | .mr(4) |
| 456 | .nr(8) |
| 457 | .kr(1) |
| 458 | .sr(1) |
| 459 | .m(4) |
| 460 | .n(8) |
| 461 | .k(8) |
| 462 | .qmax(128) |
| 463 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 464 | } |
| 465 | |
| 466 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, strided_cm) { |
| 467 | TEST_REQUIRES_ARM_NEON; |
| 468 | GemmMicrokernelTester() |
| 469 | .mr(4) |
| 470 | .nr(8) |
| 471 | .kr(1) |
| 472 | .sr(1) |
| 473 | .m(4) |
| 474 | .n(8) |
| 475 | .k(8) |
| 476 | .cm_stride(11) |
| 477 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 478 | } |
| 479 | |
| 480 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, no_a_zero_point) { |
| 481 | TEST_REQUIRES_ARM_NEON; |
| 482 | for (size_t k = 1; k <= 40; k += 9) { |
| 483 | GemmMicrokernelTester() |
| 484 | .mr(4) |
| 485 | .nr(8) |
| 486 | .kr(1) |
| 487 | .sr(1) |
| 488 | .m(4) |
| 489 | .n(8) |
| 490 | .k(k) |
| 491 | .a_zero_point(0) |
| 492 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 493 | } |
| 494 | } |
| 495 | |
| 496 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, no_b_zero_point) { |
| 497 | TEST_REQUIRES_ARM_NEON; |
| 498 | for (size_t k = 1; k <= 40; k += 9) { |
| 499 | GemmMicrokernelTester() |
| 500 | .mr(4) |
| 501 | .nr(8) |
| 502 | .kr(1) |
| 503 | .sr(1) |
| 504 | .m(4) |
| 505 | .n(8) |
| 506 | .k(k) |
| 507 | .b_zero_point(0) |
| 508 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 509 | } |
| 510 | } |
| 511 | |
| 512 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A53, no_zero_point) { |
| 513 | TEST_REQUIRES_ARM_NEON; |
| 514 | for (size_t k = 1; k <= 40; k += 9) { |
| 515 | GemmMicrokernelTester() |
| 516 | .mr(4) |
| 517 | .nr(8) |
| 518 | .kr(1) |
| 519 | .sr(1) |
| 520 | .m(4) |
| 521 | .n(8) |
| 522 | .k(k) |
| 523 | .a_zero_point(0) |
| 524 | .b_zero_point(0) |
| 525 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 526 | } |
| 527 | } |
| 528 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 529 | |
| 530 | |
| 531 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 532 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_eq_8) { |
| 533 | TEST_REQUIRES_ARM_NEON; |
| 534 | GemmMicrokernelTester() |
| 535 | .mr(4) |
| 536 | .nr(8) |
| 537 | .kr(1) |
| 538 | .sr(1) |
| 539 | .m(4) |
| 540 | .n(8) |
| 541 | .k(8) |
| 542 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 543 | } |
| 544 | |
| 545 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, strided_cn) { |
| 546 | TEST_REQUIRES_ARM_NEON; |
| 547 | GemmMicrokernelTester() |
| 548 | .mr(4) |
| 549 | .nr(8) |
| 550 | .kr(1) |
| 551 | .sr(1) |
| 552 | .m(4) |
| 553 | .n(8) |
| 554 | .k(8) |
| 555 | .cn_stride(11) |
| 556 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 557 | } |
| 558 | |
| 559 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_eq_8_strided_a) { |
| 560 | TEST_REQUIRES_ARM_NEON; |
| 561 | GemmMicrokernelTester() |
| 562 | .mr(4) |
| 563 | .nr(8) |
| 564 | .kr(1) |
| 565 | .sr(1) |
| 566 | .m(4) |
| 567 | .n(8) |
| 568 | .k(8) |
| 569 | .a_stride(11) |
| 570 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 571 | } |
| 572 | |
| 573 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_eq_8_subtile) { |
| 574 | TEST_REQUIRES_ARM_NEON; |
| 575 | for (uint32_t n = 1; n <= 8; n++) { |
| 576 | for (uint32_t m = 1; m <= 4; m++) { |
| 577 | GemmMicrokernelTester() |
| 578 | .mr(4) |
| 579 | .nr(8) |
| 580 | .kr(1) |
| 581 | .sr(1) |
| 582 | .m(m) |
| 583 | .n(n) |
| 584 | .k(8) |
| 585 | .iterations(1) |
| 586 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 587 | } |
| 588 | } |
| 589 | } |
| 590 | |
| 591 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_eq_8_subtile_m) { |
| 592 | TEST_REQUIRES_ARM_NEON; |
| 593 | for (uint32_t m = 1; m <= 4; m++) { |
| 594 | GemmMicrokernelTester() |
| 595 | .mr(4) |
| 596 | .nr(8) |
| 597 | .kr(1) |
| 598 | .sr(1) |
| 599 | .m(m) |
| 600 | .n(8) |
| 601 | .k(8) |
| 602 | .iterations(1) |
| 603 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 604 | } |
| 605 | } |
| 606 | |
| 607 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_eq_8_subtile_n) { |
| 608 | TEST_REQUIRES_ARM_NEON; |
| 609 | for (uint32_t n = 1; n <= 8; n++) { |
| 610 | GemmMicrokernelTester() |
| 611 | .mr(4) |
| 612 | .nr(8) |
| 613 | .kr(1) |
| 614 | .sr(1) |
| 615 | .m(4) |
| 616 | .n(n) |
| 617 | .k(8) |
| 618 | .iterations(1) |
| 619 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 620 | } |
| 621 | } |
| 622 | |
| 623 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_lt_8) { |
| 624 | TEST_REQUIRES_ARM_NEON; |
| 625 | for (size_t k = 1; k < 8; k++) { |
| 626 | GemmMicrokernelTester() |
| 627 | .mr(4) |
| 628 | .nr(8) |
| 629 | .kr(1) |
| 630 | .sr(1) |
| 631 | .m(4) |
| 632 | .n(8) |
| 633 | .k(k) |
| 634 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 635 | } |
| 636 | } |
| 637 | |
| 638 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_lt_8_strided_a) { |
| 639 | TEST_REQUIRES_ARM_NEON; |
| 640 | for (size_t k = 1; k < 8; k++) { |
| 641 | GemmMicrokernelTester() |
| 642 | .mr(4) |
| 643 | .nr(8) |
| 644 | .kr(1) |
| 645 | .sr(1) |
| 646 | .m(4) |
| 647 | .n(8) |
| 648 | .k(k) |
| 649 | .a_stride(11) |
| 650 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 651 | } |
| 652 | } |
| 653 | |
| 654 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_lt_8_subtile) { |
| 655 | TEST_REQUIRES_ARM_NEON; |
| 656 | for (size_t k = 1; k < 8; k++) { |
| 657 | for (uint32_t n = 1; n <= 8; n++) { |
| 658 | for (uint32_t m = 1; m <= 4; m++) { |
| 659 | GemmMicrokernelTester() |
| 660 | .mr(4) |
| 661 | .nr(8) |
| 662 | .kr(1) |
| 663 | .sr(1) |
| 664 | .m(m) |
| 665 | .n(n) |
| 666 | .k(k) |
| 667 | .iterations(1) |
| 668 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 669 | } |
| 670 | } |
| 671 | } |
| 672 | } |
| 673 | |
| 674 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_gt_8) { |
| 675 | TEST_REQUIRES_ARM_NEON; |
| 676 | for (size_t k = 9; k < 16; k++) { |
| 677 | GemmMicrokernelTester() |
| 678 | .mr(4) |
| 679 | .nr(8) |
| 680 | .kr(1) |
| 681 | .sr(1) |
| 682 | .m(4) |
| 683 | .n(8) |
| 684 | .k(k) |
| 685 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 686 | } |
| 687 | } |
| 688 | |
| 689 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_gt_8_strided_a) { |
| 690 | TEST_REQUIRES_ARM_NEON; |
| 691 | for (size_t k = 9; k < 16; k++) { |
| 692 | GemmMicrokernelTester() |
| 693 | .mr(4) |
| 694 | .nr(8) |
| 695 | .kr(1) |
| 696 | .sr(1) |
| 697 | .m(4) |
| 698 | .n(8) |
| 699 | .k(k) |
| 700 | .a_stride(19) |
| 701 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 702 | } |
| 703 | } |
| 704 | |
| 705 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_gt_8_subtile) { |
| 706 | TEST_REQUIRES_ARM_NEON; |
| 707 | for (size_t k = 9; k < 16; k++) { |
| 708 | for (uint32_t n = 1; n <= 8; n++) { |
| 709 | for (uint32_t m = 1; m <= 4; m++) { |
| 710 | GemmMicrokernelTester() |
| 711 | .mr(4) |
| 712 | .nr(8) |
| 713 | .kr(1) |
| 714 | .sr(1) |
| 715 | .m(m) |
| 716 | .n(n) |
| 717 | .k(k) |
| 718 | .iterations(1) |
| 719 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 720 | } |
| 721 | } |
| 722 | } |
| 723 | } |
| 724 | |
| 725 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_div_8) { |
| 726 | TEST_REQUIRES_ARM_NEON; |
| 727 | for (size_t k = 16; k <= 80; k += 8) { |
| 728 | GemmMicrokernelTester() |
| 729 | .mr(4) |
| 730 | .nr(8) |
| 731 | .kr(1) |
| 732 | .sr(1) |
| 733 | .m(4) |
| 734 | .n(8) |
| 735 | .k(k) |
| 736 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 737 | } |
| 738 | } |
| 739 | |
| 740 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_div_8_strided_a) { |
| 741 | TEST_REQUIRES_ARM_NEON; |
| 742 | for (size_t k = 16; k <= 80; k += 8) { |
| 743 | GemmMicrokernelTester() |
| 744 | .mr(4) |
| 745 | .nr(8) |
| 746 | .kr(1) |
| 747 | .sr(1) |
| 748 | .m(4) |
| 749 | .n(8) |
| 750 | .k(k) |
| 751 | .a_stride(83) |
| 752 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 753 | } |
| 754 | } |
| 755 | |
| 756 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, k_div_8_subtile) { |
| 757 | TEST_REQUIRES_ARM_NEON; |
| 758 | for (size_t k = 16; k <= 80; k += 8) { |
| 759 | for (uint32_t n = 1; n <= 8; n++) { |
| 760 | for (uint32_t m = 1; m <= 4; m++) { |
| 761 | GemmMicrokernelTester() |
| 762 | .mr(4) |
| 763 | .nr(8) |
| 764 | .kr(1) |
| 765 | .sr(1) |
| 766 | .m(m) |
| 767 | .n(n) |
| 768 | .k(k) |
| 769 | .iterations(1) |
| 770 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 771 | } |
| 772 | } |
| 773 | } |
| 774 | } |
| 775 | |
| 776 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_gt_8) { |
| 777 | TEST_REQUIRES_ARM_NEON; |
| 778 | for (uint32_t n = 9; n < 16; n++) { |
| 779 | for (size_t k = 1; k <= 40; k += 9) { |
| 780 | GemmMicrokernelTester() |
| 781 | .mr(4) |
| 782 | .nr(8) |
| 783 | .kr(1) |
| 784 | .sr(1) |
| 785 | .m(4) |
| 786 | .n(n) |
| 787 | .k(k) |
| 788 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 789 | } |
| 790 | } |
| 791 | } |
| 792 | |
| 793 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_gt_8_strided_cn) { |
| 794 | TEST_REQUIRES_ARM_NEON; |
| 795 | for (uint32_t n = 9; n < 16; n++) { |
| 796 | for (size_t k = 1; k <= 40; k += 9) { |
| 797 | GemmMicrokernelTester() |
| 798 | .mr(4) |
| 799 | .nr(8) |
| 800 | .kr(1) |
| 801 | .sr(1) |
| 802 | .m(4) |
| 803 | .n(n) |
| 804 | .k(k) |
| 805 | .cn_stride(11) |
| 806 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 807 | } |
| 808 | } |
| 809 | } |
| 810 | |
| 811 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_gt_8_strided_a) { |
| 812 | TEST_REQUIRES_ARM_NEON; |
| 813 | for (uint32_t n = 9; n < 16; n++) { |
| 814 | for (size_t k = 1; k <= 40; k += 9) { |
| 815 | GemmMicrokernelTester() |
| 816 | .mr(4) |
| 817 | .nr(8) |
| 818 | .kr(1) |
| 819 | .sr(1) |
| 820 | .m(4) |
| 821 | .n(n) |
| 822 | .k(k) |
| 823 | .a_stride(43) |
| 824 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 825 | } |
| 826 | } |
| 827 | } |
| 828 | |
| 829 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_gt_8_subtile) { |
| 830 | TEST_REQUIRES_ARM_NEON; |
| 831 | for (uint32_t n = 9; n < 16; n++) { |
| 832 | for (size_t k = 1; k <= 40; k += 9) { |
| 833 | for (uint32_t m = 1; m <= 4; m++) { |
| 834 | GemmMicrokernelTester() |
| 835 | .mr(4) |
| 836 | .nr(8) |
| 837 | .kr(1) |
| 838 | .sr(1) |
| 839 | .m(m) |
| 840 | .n(n) |
| 841 | .k(k) |
| 842 | .iterations(1) |
| 843 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 844 | } |
| 845 | } |
| 846 | } |
| 847 | } |
| 848 | |
| 849 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_div_8) { |
| 850 | TEST_REQUIRES_ARM_NEON; |
| 851 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 852 | for (size_t k = 1; k <= 40; k += 9) { |
| 853 | GemmMicrokernelTester() |
| 854 | .mr(4) |
| 855 | .nr(8) |
| 856 | .kr(1) |
| 857 | .sr(1) |
| 858 | .m(4) |
| 859 | .n(n) |
| 860 | .k(k) |
| 861 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 862 | } |
| 863 | } |
| 864 | } |
| 865 | |
| 866 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_div_8_strided_cn) { |
| 867 | TEST_REQUIRES_ARM_NEON; |
| 868 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 869 | for (size_t k = 1; k <= 40; k += 9) { |
| 870 | GemmMicrokernelTester() |
| 871 | .mr(4) |
| 872 | .nr(8) |
| 873 | .kr(1) |
| 874 | .sr(1) |
| 875 | .m(4) |
| 876 | .n(n) |
| 877 | .k(k) |
| 878 | .cn_stride(11) |
| 879 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 880 | } |
| 881 | } |
| 882 | } |
| 883 | |
| 884 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_div_8_strided_a) { |
| 885 | TEST_REQUIRES_ARM_NEON; |
| 886 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 887 | for (size_t k = 1; k <= 40; k += 9) { |
| 888 | GemmMicrokernelTester() |
| 889 | .mr(4) |
| 890 | .nr(8) |
| 891 | .kr(1) |
| 892 | .sr(1) |
| 893 | .m(4) |
| 894 | .n(n) |
| 895 | .k(k) |
| 896 | .a_stride(43) |
| 897 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 898 | } |
| 899 | } |
| 900 | } |
| 901 | |
| 902 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, n_div_8_subtile) { |
| 903 | TEST_REQUIRES_ARM_NEON; |
| 904 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 905 | for (size_t k = 1; k <= 40; k += 9) { |
| 906 | for (uint32_t m = 1; m <= 4; m++) { |
| 907 | GemmMicrokernelTester() |
| 908 | .mr(4) |
| 909 | .nr(8) |
| 910 | .kr(1) |
| 911 | .sr(1) |
| 912 | .m(m) |
| 913 | .n(n) |
| 914 | .k(k) |
| 915 | .iterations(1) |
| 916 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 917 | } |
| 918 | } |
| 919 | } |
| 920 | } |
| 921 | |
| 922 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, strided_cm_subtile) { |
| 923 | TEST_REQUIRES_ARM_NEON; |
| 924 | for (size_t k = 1; k <= 40; k += 9) { |
| 925 | for (uint32_t n = 1; n <= 8; n++) { |
| 926 | for (uint32_t m = 1; m <= 4; m++) { |
| 927 | GemmMicrokernelTester() |
| 928 | .mr(4) |
| 929 | .nr(8) |
| 930 | .kr(1) |
| 931 | .sr(1) |
| 932 | .m(m) |
| 933 | .n(n) |
| 934 | .k(k) |
| 935 | .cm_stride(11) |
| 936 | .iterations(1) |
| 937 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 938 | } |
| 939 | } |
| 940 | } |
| 941 | } |
| 942 | |
| 943 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, qmin) { |
| 944 | TEST_REQUIRES_ARM_NEON; |
| 945 | GemmMicrokernelTester() |
| 946 | .mr(4) |
| 947 | .nr(8) |
| 948 | .kr(1) |
| 949 | .sr(1) |
| 950 | .m(4) |
| 951 | .n(8) |
| 952 | .k(8) |
| 953 | .qmin(128) |
| 954 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 955 | } |
| 956 | |
| 957 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, qmax) { |
| 958 | TEST_REQUIRES_ARM_NEON; |
| 959 | GemmMicrokernelTester() |
| 960 | .mr(4) |
| 961 | .nr(8) |
| 962 | .kr(1) |
| 963 | .sr(1) |
| 964 | .m(4) |
| 965 | .n(8) |
| 966 | .k(8) |
| 967 | .qmax(128) |
| 968 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 969 | } |
| 970 | |
| 971 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, strided_cm) { |
| 972 | TEST_REQUIRES_ARM_NEON; |
| 973 | GemmMicrokernelTester() |
| 974 | .mr(4) |
| 975 | .nr(8) |
| 976 | .kr(1) |
| 977 | .sr(1) |
| 978 | .m(4) |
| 979 | .n(8) |
| 980 | .k(8) |
| 981 | .cm_stride(11) |
| 982 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 983 | } |
| 984 | |
| 985 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, no_a_zero_point) { |
| 986 | TEST_REQUIRES_ARM_NEON; |
| 987 | for (size_t k = 1; k <= 40; k += 9) { |
| 988 | GemmMicrokernelTester() |
| 989 | .mr(4) |
| 990 | .nr(8) |
| 991 | .kr(1) |
| 992 | .sr(1) |
| 993 | .m(4) |
| 994 | .n(8) |
| 995 | .k(k) |
| 996 | .a_zero_point(0) |
| 997 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 998 | } |
| 999 | } |
| 1000 | |
| 1001 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, no_b_zero_point) { |
| 1002 | TEST_REQUIRES_ARM_NEON; |
| 1003 | for (size_t k = 1; k <= 40; k += 9) { |
| 1004 | GemmMicrokernelTester() |
| 1005 | .mr(4) |
| 1006 | .nr(8) |
| 1007 | .kr(1) |
| 1008 | .sr(1) |
| 1009 | .m(4) |
| 1010 | .n(8) |
| 1011 | .k(k) |
| 1012 | .b_zero_point(0) |
| 1013 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1014 | } |
| 1015 | } |
| 1016 | |
| 1017 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A53, no_zero_point) { |
| 1018 | TEST_REQUIRES_ARM_NEON; |
| 1019 | for (size_t k = 1; k <= 40; k += 9) { |
| 1020 | GemmMicrokernelTester() |
| 1021 | .mr(4) |
| 1022 | .nr(8) |
| 1023 | .kr(1) |
| 1024 | .sr(1) |
| 1025 | .m(4) |
| 1026 | .n(8) |
| 1027 | .k(k) |
| 1028 | .a_zero_point(0) |
| 1029 | .b_zero_point(0) |
| 1030 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1031 | } |
| 1032 | } |
| 1033 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 1034 | |
| 1035 | |
Frank Barchard | 34251d8 | 2022-02-02 11:57:11 -0800 | [diff] [blame] | 1036 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 1037 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_eq_8) { |
| 1038 | TEST_REQUIRES_ARM_NEON; |
| 1039 | GemmMicrokernelTester() |
| 1040 | .mr(4) |
| 1041 | .nr(8) |
| 1042 | .kr(1) |
| 1043 | .sr(1) |
| 1044 | .m(4) |
| 1045 | .n(8) |
| 1046 | .k(8) |
| 1047 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1048 | } |
| 1049 | |
| 1050 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, strided_cn) { |
| 1051 | TEST_REQUIRES_ARM_NEON; |
| 1052 | GemmMicrokernelTester() |
| 1053 | .mr(4) |
| 1054 | .nr(8) |
| 1055 | .kr(1) |
| 1056 | .sr(1) |
| 1057 | .m(4) |
| 1058 | .n(8) |
| 1059 | .k(8) |
| 1060 | .cn_stride(11) |
| 1061 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1062 | } |
| 1063 | |
| 1064 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_eq_8_strided_a) { |
| 1065 | TEST_REQUIRES_ARM_NEON; |
| 1066 | GemmMicrokernelTester() |
| 1067 | .mr(4) |
| 1068 | .nr(8) |
| 1069 | .kr(1) |
| 1070 | .sr(1) |
| 1071 | .m(4) |
| 1072 | .n(8) |
| 1073 | .k(8) |
| 1074 | .a_stride(11) |
| 1075 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1076 | } |
| 1077 | |
| 1078 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_eq_8_subtile) { |
| 1079 | TEST_REQUIRES_ARM_NEON; |
| 1080 | for (uint32_t n = 1; n <= 8; n++) { |
| 1081 | for (uint32_t m = 1; m <= 4; m++) { |
| 1082 | GemmMicrokernelTester() |
| 1083 | .mr(4) |
| 1084 | .nr(8) |
| 1085 | .kr(1) |
| 1086 | .sr(1) |
| 1087 | .m(m) |
| 1088 | .n(n) |
| 1089 | .k(8) |
| 1090 | .iterations(1) |
| 1091 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1092 | } |
| 1093 | } |
| 1094 | } |
| 1095 | |
| 1096 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_eq_8_subtile_m) { |
| 1097 | TEST_REQUIRES_ARM_NEON; |
| 1098 | for (uint32_t m = 1; m <= 4; m++) { |
| 1099 | GemmMicrokernelTester() |
| 1100 | .mr(4) |
| 1101 | .nr(8) |
| 1102 | .kr(1) |
| 1103 | .sr(1) |
| 1104 | .m(m) |
| 1105 | .n(8) |
| 1106 | .k(8) |
| 1107 | .iterations(1) |
| 1108 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1109 | } |
| 1110 | } |
| 1111 | |
| 1112 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_eq_8_subtile_n) { |
| 1113 | TEST_REQUIRES_ARM_NEON; |
| 1114 | for (uint32_t n = 1; n <= 8; n++) { |
| 1115 | GemmMicrokernelTester() |
| 1116 | .mr(4) |
| 1117 | .nr(8) |
| 1118 | .kr(1) |
| 1119 | .sr(1) |
| 1120 | .m(4) |
| 1121 | .n(n) |
| 1122 | .k(8) |
| 1123 | .iterations(1) |
| 1124 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1125 | } |
| 1126 | } |
| 1127 | |
| 1128 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_lt_8) { |
| 1129 | TEST_REQUIRES_ARM_NEON; |
| 1130 | for (size_t k = 1; k < 8; k++) { |
| 1131 | GemmMicrokernelTester() |
| 1132 | .mr(4) |
| 1133 | .nr(8) |
| 1134 | .kr(1) |
| 1135 | .sr(1) |
| 1136 | .m(4) |
| 1137 | .n(8) |
| 1138 | .k(k) |
| 1139 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1140 | } |
| 1141 | } |
| 1142 | |
| 1143 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_lt_8_strided_a) { |
| 1144 | TEST_REQUIRES_ARM_NEON; |
| 1145 | for (size_t k = 1; k < 8; k++) { |
| 1146 | GemmMicrokernelTester() |
| 1147 | .mr(4) |
| 1148 | .nr(8) |
| 1149 | .kr(1) |
| 1150 | .sr(1) |
| 1151 | .m(4) |
| 1152 | .n(8) |
| 1153 | .k(k) |
| 1154 | .a_stride(11) |
| 1155 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1156 | } |
| 1157 | } |
| 1158 | |
| 1159 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_lt_8_subtile) { |
| 1160 | TEST_REQUIRES_ARM_NEON; |
| 1161 | for (size_t k = 1; k < 8; k++) { |
| 1162 | for (uint32_t n = 1; n <= 8; n++) { |
| 1163 | for (uint32_t m = 1; m <= 4; m++) { |
| 1164 | GemmMicrokernelTester() |
| 1165 | .mr(4) |
| 1166 | .nr(8) |
| 1167 | .kr(1) |
| 1168 | .sr(1) |
| 1169 | .m(m) |
| 1170 | .n(n) |
| 1171 | .k(k) |
| 1172 | .iterations(1) |
| 1173 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1174 | } |
| 1175 | } |
| 1176 | } |
| 1177 | } |
| 1178 | |
| 1179 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_gt_8) { |
| 1180 | TEST_REQUIRES_ARM_NEON; |
| 1181 | for (size_t k = 9; k < 16; k++) { |
| 1182 | GemmMicrokernelTester() |
| 1183 | .mr(4) |
| 1184 | .nr(8) |
| 1185 | .kr(1) |
| 1186 | .sr(1) |
| 1187 | .m(4) |
| 1188 | .n(8) |
| 1189 | .k(k) |
| 1190 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1191 | } |
| 1192 | } |
| 1193 | |
| 1194 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_gt_8_strided_a) { |
| 1195 | TEST_REQUIRES_ARM_NEON; |
| 1196 | for (size_t k = 9; k < 16; k++) { |
| 1197 | GemmMicrokernelTester() |
| 1198 | .mr(4) |
| 1199 | .nr(8) |
| 1200 | .kr(1) |
| 1201 | .sr(1) |
| 1202 | .m(4) |
| 1203 | .n(8) |
| 1204 | .k(k) |
| 1205 | .a_stride(19) |
| 1206 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1207 | } |
| 1208 | } |
| 1209 | |
| 1210 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_gt_8_subtile) { |
| 1211 | TEST_REQUIRES_ARM_NEON; |
| 1212 | for (size_t k = 9; k < 16; k++) { |
| 1213 | for (uint32_t n = 1; n <= 8; n++) { |
| 1214 | for (uint32_t m = 1; m <= 4; m++) { |
| 1215 | GemmMicrokernelTester() |
| 1216 | .mr(4) |
| 1217 | .nr(8) |
| 1218 | .kr(1) |
| 1219 | .sr(1) |
| 1220 | .m(m) |
| 1221 | .n(n) |
| 1222 | .k(k) |
| 1223 | .iterations(1) |
| 1224 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1225 | } |
| 1226 | } |
| 1227 | } |
| 1228 | } |
| 1229 | |
| 1230 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_div_8) { |
| 1231 | TEST_REQUIRES_ARM_NEON; |
| 1232 | for (size_t k = 16; k <= 80; k += 8) { |
| 1233 | GemmMicrokernelTester() |
| 1234 | .mr(4) |
| 1235 | .nr(8) |
| 1236 | .kr(1) |
| 1237 | .sr(1) |
| 1238 | .m(4) |
| 1239 | .n(8) |
| 1240 | .k(k) |
| 1241 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1242 | } |
| 1243 | } |
| 1244 | |
| 1245 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_div_8_strided_a) { |
| 1246 | TEST_REQUIRES_ARM_NEON; |
| 1247 | for (size_t k = 16; k <= 80; k += 8) { |
| 1248 | GemmMicrokernelTester() |
| 1249 | .mr(4) |
| 1250 | .nr(8) |
| 1251 | .kr(1) |
| 1252 | .sr(1) |
| 1253 | .m(4) |
| 1254 | .n(8) |
| 1255 | .k(k) |
| 1256 | .a_stride(83) |
| 1257 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1258 | } |
| 1259 | } |
| 1260 | |
| 1261 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, k_div_8_subtile) { |
| 1262 | TEST_REQUIRES_ARM_NEON; |
| 1263 | for (size_t k = 16; k <= 80; k += 8) { |
| 1264 | for (uint32_t n = 1; n <= 8; n++) { |
| 1265 | for (uint32_t m = 1; m <= 4; m++) { |
| 1266 | GemmMicrokernelTester() |
| 1267 | .mr(4) |
| 1268 | .nr(8) |
| 1269 | .kr(1) |
| 1270 | .sr(1) |
| 1271 | .m(m) |
| 1272 | .n(n) |
| 1273 | .k(k) |
| 1274 | .iterations(1) |
| 1275 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1276 | } |
| 1277 | } |
| 1278 | } |
| 1279 | } |
| 1280 | |
| 1281 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_gt_8) { |
| 1282 | TEST_REQUIRES_ARM_NEON; |
| 1283 | for (uint32_t n = 9; n < 16; n++) { |
| 1284 | for (size_t k = 1; k <= 40; k += 9) { |
| 1285 | GemmMicrokernelTester() |
| 1286 | .mr(4) |
| 1287 | .nr(8) |
| 1288 | .kr(1) |
| 1289 | .sr(1) |
| 1290 | .m(4) |
| 1291 | .n(n) |
| 1292 | .k(k) |
| 1293 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1294 | } |
| 1295 | } |
| 1296 | } |
| 1297 | |
| 1298 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_gt_8_strided_cn) { |
| 1299 | TEST_REQUIRES_ARM_NEON; |
| 1300 | for (uint32_t n = 9; n < 16; n++) { |
| 1301 | for (size_t k = 1; k <= 40; k += 9) { |
| 1302 | GemmMicrokernelTester() |
| 1303 | .mr(4) |
| 1304 | .nr(8) |
| 1305 | .kr(1) |
| 1306 | .sr(1) |
| 1307 | .m(4) |
| 1308 | .n(n) |
| 1309 | .k(k) |
| 1310 | .cn_stride(11) |
| 1311 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1312 | } |
| 1313 | } |
| 1314 | } |
| 1315 | |
| 1316 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_gt_8_strided_a) { |
| 1317 | TEST_REQUIRES_ARM_NEON; |
| 1318 | for (uint32_t n = 9; n < 16; n++) { |
| 1319 | for (size_t k = 1; k <= 40; k += 9) { |
| 1320 | GemmMicrokernelTester() |
| 1321 | .mr(4) |
| 1322 | .nr(8) |
| 1323 | .kr(1) |
| 1324 | .sr(1) |
| 1325 | .m(4) |
| 1326 | .n(n) |
| 1327 | .k(k) |
| 1328 | .a_stride(43) |
| 1329 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1330 | } |
| 1331 | } |
| 1332 | } |
| 1333 | |
| 1334 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_gt_8_subtile) { |
| 1335 | TEST_REQUIRES_ARM_NEON; |
| 1336 | for (uint32_t n = 9; n < 16; n++) { |
| 1337 | for (size_t k = 1; k <= 40; k += 9) { |
| 1338 | for (uint32_t m = 1; m <= 4; m++) { |
| 1339 | GemmMicrokernelTester() |
| 1340 | .mr(4) |
| 1341 | .nr(8) |
| 1342 | .kr(1) |
| 1343 | .sr(1) |
| 1344 | .m(m) |
| 1345 | .n(n) |
| 1346 | .k(k) |
| 1347 | .iterations(1) |
| 1348 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1349 | } |
| 1350 | } |
| 1351 | } |
| 1352 | } |
| 1353 | |
| 1354 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_div_8) { |
| 1355 | TEST_REQUIRES_ARM_NEON; |
| 1356 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1357 | for (size_t k = 1; k <= 40; k += 9) { |
| 1358 | GemmMicrokernelTester() |
| 1359 | .mr(4) |
| 1360 | .nr(8) |
| 1361 | .kr(1) |
| 1362 | .sr(1) |
| 1363 | .m(4) |
| 1364 | .n(n) |
| 1365 | .k(k) |
| 1366 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1367 | } |
| 1368 | } |
| 1369 | } |
| 1370 | |
| 1371 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_div_8_strided_cn) { |
| 1372 | TEST_REQUIRES_ARM_NEON; |
| 1373 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1374 | for (size_t k = 1; k <= 40; k += 9) { |
| 1375 | GemmMicrokernelTester() |
| 1376 | .mr(4) |
| 1377 | .nr(8) |
| 1378 | .kr(1) |
| 1379 | .sr(1) |
| 1380 | .m(4) |
| 1381 | .n(n) |
| 1382 | .k(k) |
| 1383 | .cn_stride(11) |
| 1384 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1385 | } |
| 1386 | } |
| 1387 | } |
| 1388 | |
| 1389 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_div_8_strided_a) { |
| 1390 | TEST_REQUIRES_ARM_NEON; |
| 1391 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1392 | for (size_t k = 1; k <= 40; k += 9) { |
| 1393 | GemmMicrokernelTester() |
| 1394 | .mr(4) |
| 1395 | .nr(8) |
| 1396 | .kr(1) |
| 1397 | .sr(1) |
| 1398 | .m(4) |
| 1399 | .n(n) |
| 1400 | .k(k) |
| 1401 | .a_stride(43) |
| 1402 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1403 | } |
| 1404 | } |
| 1405 | } |
| 1406 | |
| 1407 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, n_div_8_subtile) { |
| 1408 | TEST_REQUIRES_ARM_NEON; |
| 1409 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1410 | for (size_t k = 1; k <= 40; k += 9) { |
| 1411 | for (uint32_t m = 1; m <= 4; m++) { |
| 1412 | GemmMicrokernelTester() |
| 1413 | .mr(4) |
| 1414 | .nr(8) |
| 1415 | .kr(1) |
| 1416 | .sr(1) |
| 1417 | .m(m) |
| 1418 | .n(n) |
| 1419 | .k(k) |
| 1420 | .iterations(1) |
| 1421 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1422 | } |
| 1423 | } |
| 1424 | } |
| 1425 | } |
| 1426 | |
| 1427 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, strided_cm_subtile) { |
| 1428 | TEST_REQUIRES_ARM_NEON; |
| 1429 | for (size_t k = 1; k <= 40; k += 9) { |
| 1430 | for (uint32_t n = 1; n <= 8; n++) { |
| 1431 | for (uint32_t m = 1; m <= 4; m++) { |
| 1432 | GemmMicrokernelTester() |
| 1433 | .mr(4) |
| 1434 | .nr(8) |
| 1435 | .kr(1) |
| 1436 | .sr(1) |
| 1437 | .m(m) |
| 1438 | .n(n) |
| 1439 | .k(k) |
| 1440 | .cm_stride(11) |
| 1441 | .iterations(1) |
| 1442 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1443 | } |
| 1444 | } |
| 1445 | } |
| 1446 | } |
| 1447 | |
| 1448 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, qmin) { |
| 1449 | TEST_REQUIRES_ARM_NEON; |
| 1450 | GemmMicrokernelTester() |
| 1451 | .mr(4) |
| 1452 | .nr(8) |
| 1453 | .kr(1) |
| 1454 | .sr(1) |
| 1455 | .m(4) |
| 1456 | .n(8) |
| 1457 | .k(8) |
| 1458 | .qmin(128) |
| 1459 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1460 | } |
| 1461 | |
| 1462 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, qmax) { |
| 1463 | TEST_REQUIRES_ARM_NEON; |
| 1464 | GemmMicrokernelTester() |
| 1465 | .mr(4) |
| 1466 | .nr(8) |
| 1467 | .kr(1) |
| 1468 | .sr(1) |
| 1469 | .m(4) |
| 1470 | .n(8) |
| 1471 | .k(8) |
| 1472 | .qmax(128) |
| 1473 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1474 | } |
| 1475 | |
| 1476 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, strided_cm) { |
| 1477 | TEST_REQUIRES_ARM_NEON; |
| 1478 | GemmMicrokernelTester() |
| 1479 | .mr(4) |
| 1480 | .nr(8) |
| 1481 | .kr(1) |
| 1482 | .sr(1) |
| 1483 | .m(4) |
| 1484 | .n(8) |
| 1485 | .k(8) |
| 1486 | .cm_stride(11) |
| 1487 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1488 | } |
| 1489 | |
| 1490 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, no_a_zero_point) { |
| 1491 | TEST_REQUIRES_ARM_NEON; |
| 1492 | for (size_t k = 1; k <= 40; k += 9) { |
| 1493 | GemmMicrokernelTester() |
| 1494 | .mr(4) |
| 1495 | .nr(8) |
| 1496 | .kr(1) |
| 1497 | .sr(1) |
| 1498 | .m(4) |
| 1499 | .n(8) |
| 1500 | .k(k) |
| 1501 | .a_zero_point(0) |
| 1502 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1503 | } |
| 1504 | } |
| 1505 | |
| 1506 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, no_b_zero_point) { |
| 1507 | TEST_REQUIRES_ARM_NEON; |
| 1508 | for (size_t k = 1; k <= 40; k += 9) { |
| 1509 | GemmMicrokernelTester() |
| 1510 | .mr(4) |
| 1511 | .nr(8) |
| 1512 | .kr(1) |
| 1513 | .sr(1) |
| 1514 | .m(4) |
| 1515 | .n(8) |
| 1516 | .k(k) |
| 1517 | .b_zero_point(0) |
| 1518 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1519 | } |
| 1520 | } |
| 1521 | |
| 1522 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_CORTEX_A7, no_zero_point) { |
| 1523 | TEST_REQUIRES_ARM_NEON; |
| 1524 | for (size_t k = 1; k <= 40; k += 9) { |
| 1525 | GemmMicrokernelTester() |
| 1526 | .mr(4) |
| 1527 | .nr(8) |
| 1528 | .kr(1) |
| 1529 | .sr(1) |
| 1530 | .m(4) |
| 1531 | .n(8) |
| 1532 | .k(k) |
| 1533 | .a_zero_point(0) |
| 1534 | .b_zero_point(0) |
| 1535 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1536 | } |
| 1537 | } |
| 1538 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 1539 | |
| 1540 | |
| 1541 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 1542 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_eq_8) { |
| 1543 | TEST_REQUIRES_ARM_NEON; |
| 1544 | GemmMicrokernelTester() |
| 1545 | .mr(4) |
| 1546 | .nr(8) |
| 1547 | .kr(1) |
| 1548 | .sr(1) |
| 1549 | .m(4) |
| 1550 | .n(8) |
| 1551 | .k(8) |
| 1552 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1553 | } |
| 1554 | |
| 1555 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, strided_cn) { |
| 1556 | TEST_REQUIRES_ARM_NEON; |
| 1557 | GemmMicrokernelTester() |
| 1558 | .mr(4) |
| 1559 | .nr(8) |
| 1560 | .kr(1) |
| 1561 | .sr(1) |
| 1562 | .m(4) |
| 1563 | .n(8) |
| 1564 | .k(8) |
| 1565 | .cn_stride(11) |
| 1566 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1567 | } |
| 1568 | |
| 1569 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_eq_8_strided_a) { |
| 1570 | TEST_REQUIRES_ARM_NEON; |
| 1571 | GemmMicrokernelTester() |
| 1572 | .mr(4) |
| 1573 | .nr(8) |
| 1574 | .kr(1) |
| 1575 | .sr(1) |
| 1576 | .m(4) |
| 1577 | .n(8) |
| 1578 | .k(8) |
| 1579 | .a_stride(11) |
| 1580 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1581 | } |
| 1582 | |
| 1583 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_eq_8_subtile) { |
| 1584 | TEST_REQUIRES_ARM_NEON; |
| 1585 | for (uint32_t n = 1; n <= 8; n++) { |
| 1586 | for (uint32_t m = 1; m <= 4; m++) { |
| 1587 | GemmMicrokernelTester() |
| 1588 | .mr(4) |
| 1589 | .nr(8) |
| 1590 | .kr(1) |
| 1591 | .sr(1) |
| 1592 | .m(m) |
| 1593 | .n(n) |
| 1594 | .k(8) |
| 1595 | .iterations(1) |
| 1596 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1597 | } |
| 1598 | } |
| 1599 | } |
| 1600 | |
| 1601 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_eq_8_subtile_m) { |
| 1602 | TEST_REQUIRES_ARM_NEON; |
| 1603 | for (uint32_t m = 1; m <= 4; m++) { |
| 1604 | GemmMicrokernelTester() |
| 1605 | .mr(4) |
| 1606 | .nr(8) |
| 1607 | .kr(1) |
| 1608 | .sr(1) |
| 1609 | .m(m) |
| 1610 | .n(8) |
| 1611 | .k(8) |
| 1612 | .iterations(1) |
| 1613 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1614 | } |
| 1615 | } |
| 1616 | |
| 1617 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_eq_8_subtile_n) { |
| 1618 | TEST_REQUIRES_ARM_NEON; |
| 1619 | for (uint32_t n = 1; n <= 8; n++) { |
| 1620 | GemmMicrokernelTester() |
| 1621 | .mr(4) |
| 1622 | .nr(8) |
| 1623 | .kr(1) |
| 1624 | .sr(1) |
| 1625 | .m(4) |
| 1626 | .n(n) |
| 1627 | .k(8) |
| 1628 | .iterations(1) |
| 1629 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1630 | } |
| 1631 | } |
| 1632 | |
| 1633 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_lt_8) { |
| 1634 | TEST_REQUIRES_ARM_NEON; |
| 1635 | for (size_t k = 1; k < 8; k++) { |
| 1636 | GemmMicrokernelTester() |
| 1637 | .mr(4) |
| 1638 | .nr(8) |
| 1639 | .kr(1) |
| 1640 | .sr(1) |
| 1641 | .m(4) |
| 1642 | .n(8) |
| 1643 | .k(k) |
| 1644 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1645 | } |
| 1646 | } |
| 1647 | |
| 1648 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_lt_8_strided_a) { |
| 1649 | TEST_REQUIRES_ARM_NEON; |
| 1650 | for (size_t k = 1; k < 8; k++) { |
| 1651 | GemmMicrokernelTester() |
| 1652 | .mr(4) |
| 1653 | .nr(8) |
| 1654 | .kr(1) |
| 1655 | .sr(1) |
| 1656 | .m(4) |
| 1657 | .n(8) |
| 1658 | .k(k) |
| 1659 | .a_stride(11) |
| 1660 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1661 | } |
| 1662 | } |
| 1663 | |
| 1664 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_lt_8_subtile) { |
| 1665 | TEST_REQUIRES_ARM_NEON; |
| 1666 | for (size_t k = 1; k < 8; k++) { |
| 1667 | for (uint32_t n = 1; n <= 8; n++) { |
| 1668 | for (uint32_t m = 1; m <= 4; m++) { |
| 1669 | GemmMicrokernelTester() |
| 1670 | .mr(4) |
| 1671 | .nr(8) |
| 1672 | .kr(1) |
| 1673 | .sr(1) |
| 1674 | .m(m) |
| 1675 | .n(n) |
| 1676 | .k(k) |
| 1677 | .iterations(1) |
| 1678 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1679 | } |
| 1680 | } |
| 1681 | } |
| 1682 | } |
| 1683 | |
| 1684 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_gt_8) { |
| 1685 | TEST_REQUIRES_ARM_NEON; |
| 1686 | for (size_t k = 9; k < 16; k++) { |
| 1687 | GemmMicrokernelTester() |
| 1688 | .mr(4) |
| 1689 | .nr(8) |
| 1690 | .kr(1) |
| 1691 | .sr(1) |
| 1692 | .m(4) |
| 1693 | .n(8) |
| 1694 | .k(k) |
| 1695 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1696 | } |
| 1697 | } |
| 1698 | |
| 1699 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_gt_8_strided_a) { |
| 1700 | TEST_REQUIRES_ARM_NEON; |
| 1701 | for (size_t k = 9; k < 16; k++) { |
| 1702 | GemmMicrokernelTester() |
| 1703 | .mr(4) |
| 1704 | .nr(8) |
| 1705 | .kr(1) |
| 1706 | .sr(1) |
| 1707 | .m(4) |
| 1708 | .n(8) |
| 1709 | .k(k) |
| 1710 | .a_stride(19) |
| 1711 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1712 | } |
| 1713 | } |
| 1714 | |
| 1715 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_gt_8_subtile) { |
| 1716 | TEST_REQUIRES_ARM_NEON; |
| 1717 | for (size_t k = 9; k < 16; k++) { |
| 1718 | for (uint32_t n = 1; n <= 8; n++) { |
| 1719 | for (uint32_t m = 1; m <= 4; m++) { |
| 1720 | GemmMicrokernelTester() |
| 1721 | .mr(4) |
| 1722 | .nr(8) |
| 1723 | .kr(1) |
| 1724 | .sr(1) |
| 1725 | .m(m) |
| 1726 | .n(n) |
| 1727 | .k(k) |
| 1728 | .iterations(1) |
| 1729 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1730 | } |
| 1731 | } |
| 1732 | } |
| 1733 | } |
| 1734 | |
| 1735 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_div_8) { |
| 1736 | TEST_REQUIRES_ARM_NEON; |
| 1737 | for (size_t k = 16; k <= 80; k += 8) { |
| 1738 | GemmMicrokernelTester() |
| 1739 | .mr(4) |
| 1740 | .nr(8) |
| 1741 | .kr(1) |
| 1742 | .sr(1) |
| 1743 | .m(4) |
| 1744 | .n(8) |
| 1745 | .k(k) |
| 1746 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1747 | } |
| 1748 | } |
| 1749 | |
| 1750 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_div_8_strided_a) { |
| 1751 | TEST_REQUIRES_ARM_NEON; |
| 1752 | for (size_t k = 16; k <= 80; k += 8) { |
| 1753 | GemmMicrokernelTester() |
| 1754 | .mr(4) |
| 1755 | .nr(8) |
| 1756 | .kr(1) |
| 1757 | .sr(1) |
| 1758 | .m(4) |
| 1759 | .n(8) |
| 1760 | .k(k) |
| 1761 | .a_stride(83) |
| 1762 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1763 | } |
| 1764 | } |
| 1765 | |
| 1766 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, k_div_8_subtile) { |
| 1767 | TEST_REQUIRES_ARM_NEON; |
| 1768 | for (size_t k = 16; k <= 80; k += 8) { |
| 1769 | for (uint32_t n = 1; n <= 8; n++) { |
| 1770 | for (uint32_t m = 1; m <= 4; m++) { |
| 1771 | GemmMicrokernelTester() |
| 1772 | .mr(4) |
| 1773 | .nr(8) |
| 1774 | .kr(1) |
| 1775 | .sr(1) |
| 1776 | .m(m) |
| 1777 | .n(n) |
| 1778 | .k(k) |
| 1779 | .iterations(1) |
| 1780 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1781 | } |
| 1782 | } |
| 1783 | } |
| 1784 | } |
| 1785 | |
| 1786 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_gt_8) { |
| 1787 | TEST_REQUIRES_ARM_NEON; |
| 1788 | for (uint32_t n = 9; n < 16; n++) { |
| 1789 | for (size_t k = 1; k <= 40; k += 9) { |
| 1790 | GemmMicrokernelTester() |
| 1791 | .mr(4) |
| 1792 | .nr(8) |
| 1793 | .kr(1) |
| 1794 | .sr(1) |
| 1795 | .m(4) |
| 1796 | .n(n) |
| 1797 | .k(k) |
| 1798 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1799 | } |
| 1800 | } |
| 1801 | } |
| 1802 | |
| 1803 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_gt_8_strided_cn) { |
| 1804 | TEST_REQUIRES_ARM_NEON; |
| 1805 | for (uint32_t n = 9; n < 16; n++) { |
| 1806 | for (size_t k = 1; k <= 40; k += 9) { |
| 1807 | GemmMicrokernelTester() |
| 1808 | .mr(4) |
| 1809 | .nr(8) |
| 1810 | .kr(1) |
| 1811 | .sr(1) |
| 1812 | .m(4) |
| 1813 | .n(n) |
| 1814 | .k(k) |
| 1815 | .cn_stride(11) |
| 1816 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1817 | } |
| 1818 | } |
| 1819 | } |
| 1820 | |
| 1821 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_gt_8_strided_a) { |
| 1822 | TEST_REQUIRES_ARM_NEON; |
| 1823 | for (uint32_t n = 9; n < 16; n++) { |
| 1824 | for (size_t k = 1; k <= 40; k += 9) { |
| 1825 | GemmMicrokernelTester() |
| 1826 | .mr(4) |
| 1827 | .nr(8) |
| 1828 | .kr(1) |
| 1829 | .sr(1) |
| 1830 | .m(4) |
| 1831 | .n(n) |
| 1832 | .k(k) |
| 1833 | .a_stride(43) |
| 1834 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1835 | } |
| 1836 | } |
| 1837 | } |
| 1838 | |
| 1839 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_gt_8_subtile) { |
| 1840 | TEST_REQUIRES_ARM_NEON; |
| 1841 | for (uint32_t n = 9; n < 16; n++) { |
| 1842 | for (size_t k = 1; k <= 40; k += 9) { |
| 1843 | for (uint32_t m = 1; m <= 4; m++) { |
| 1844 | GemmMicrokernelTester() |
| 1845 | .mr(4) |
| 1846 | .nr(8) |
| 1847 | .kr(1) |
| 1848 | .sr(1) |
| 1849 | .m(m) |
| 1850 | .n(n) |
| 1851 | .k(k) |
| 1852 | .iterations(1) |
| 1853 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1854 | } |
| 1855 | } |
| 1856 | } |
| 1857 | } |
| 1858 | |
| 1859 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_div_8) { |
| 1860 | TEST_REQUIRES_ARM_NEON; |
| 1861 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1862 | for (size_t k = 1; k <= 40; k += 9) { |
| 1863 | GemmMicrokernelTester() |
| 1864 | .mr(4) |
| 1865 | .nr(8) |
| 1866 | .kr(1) |
| 1867 | .sr(1) |
| 1868 | .m(4) |
| 1869 | .n(n) |
| 1870 | .k(k) |
| 1871 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1872 | } |
| 1873 | } |
| 1874 | } |
| 1875 | |
| 1876 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_div_8_strided_cn) { |
| 1877 | TEST_REQUIRES_ARM_NEON; |
| 1878 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1879 | for (size_t k = 1; k <= 40; k += 9) { |
| 1880 | GemmMicrokernelTester() |
| 1881 | .mr(4) |
| 1882 | .nr(8) |
| 1883 | .kr(1) |
| 1884 | .sr(1) |
| 1885 | .m(4) |
| 1886 | .n(n) |
| 1887 | .k(k) |
| 1888 | .cn_stride(11) |
| 1889 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1890 | } |
| 1891 | } |
| 1892 | } |
| 1893 | |
| 1894 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_div_8_strided_a) { |
| 1895 | TEST_REQUIRES_ARM_NEON; |
| 1896 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1897 | for (size_t k = 1; k <= 40; k += 9) { |
| 1898 | GemmMicrokernelTester() |
| 1899 | .mr(4) |
| 1900 | .nr(8) |
| 1901 | .kr(1) |
| 1902 | .sr(1) |
| 1903 | .m(4) |
| 1904 | .n(n) |
| 1905 | .k(k) |
| 1906 | .a_stride(43) |
| 1907 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1908 | } |
| 1909 | } |
| 1910 | } |
| 1911 | |
| 1912 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, n_div_8_subtile) { |
| 1913 | TEST_REQUIRES_ARM_NEON; |
| 1914 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1915 | for (size_t k = 1; k <= 40; k += 9) { |
| 1916 | for (uint32_t m = 1; m <= 4; m++) { |
| 1917 | GemmMicrokernelTester() |
| 1918 | .mr(4) |
| 1919 | .nr(8) |
| 1920 | .kr(1) |
| 1921 | .sr(1) |
| 1922 | .m(m) |
| 1923 | .n(n) |
| 1924 | .k(k) |
| 1925 | .iterations(1) |
| 1926 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1927 | } |
| 1928 | } |
| 1929 | } |
| 1930 | } |
| 1931 | |
| 1932 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, strided_cm_subtile) { |
| 1933 | TEST_REQUIRES_ARM_NEON; |
| 1934 | for (size_t k = 1; k <= 40; k += 9) { |
| 1935 | for (uint32_t n = 1; n <= 8; n++) { |
| 1936 | for (uint32_t m = 1; m <= 4; m++) { |
| 1937 | GemmMicrokernelTester() |
| 1938 | .mr(4) |
| 1939 | .nr(8) |
| 1940 | .kr(1) |
| 1941 | .sr(1) |
| 1942 | .m(m) |
| 1943 | .n(n) |
| 1944 | .k(k) |
| 1945 | .cm_stride(11) |
| 1946 | .iterations(1) |
| 1947 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1948 | } |
| 1949 | } |
| 1950 | } |
| 1951 | } |
| 1952 | |
| 1953 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, qmin) { |
| 1954 | TEST_REQUIRES_ARM_NEON; |
| 1955 | GemmMicrokernelTester() |
| 1956 | .mr(4) |
| 1957 | .nr(8) |
| 1958 | .kr(1) |
| 1959 | .sr(1) |
| 1960 | .m(4) |
| 1961 | .n(8) |
| 1962 | .k(8) |
| 1963 | .qmin(128) |
| 1964 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1965 | } |
| 1966 | |
| 1967 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, qmax) { |
| 1968 | TEST_REQUIRES_ARM_NEON; |
| 1969 | GemmMicrokernelTester() |
| 1970 | .mr(4) |
| 1971 | .nr(8) |
| 1972 | .kr(1) |
| 1973 | .sr(1) |
| 1974 | .m(4) |
| 1975 | .n(8) |
| 1976 | .k(8) |
| 1977 | .qmax(128) |
| 1978 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1979 | } |
| 1980 | |
| 1981 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, strided_cm) { |
| 1982 | TEST_REQUIRES_ARM_NEON; |
| 1983 | GemmMicrokernelTester() |
| 1984 | .mr(4) |
| 1985 | .nr(8) |
| 1986 | .kr(1) |
| 1987 | .sr(1) |
| 1988 | .m(4) |
| 1989 | .n(8) |
| 1990 | .k(8) |
| 1991 | .cm_stride(11) |
| 1992 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 1993 | } |
| 1994 | |
| 1995 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, no_a_zero_point) { |
| 1996 | TEST_REQUIRES_ARM_NEON; |
| 1997 | for (size_t k = 1; k <= 40; k += 9) { |
| 1998 | GemmMicrokernelTester() |
| 1999 | .mr(4) |
| 2000 | .nr(8) |
| 2001 | .kr(1) |
| 2002 | .sr(1) |
| 2003 | .m(4) |
| 2004 | .n(8) |
| 2005 | .k(k) |
| 2006 | .a_zero_point(0) |
| 2007 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 2008 | } |
| 2009 | } |
| 2010 | |
| 2011 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, no_b_zero_point) { |
| 2012 | TEST_REQUIRES_ARM_NEON; |
| 2013 | for (size_t k = 1; k <= 40; k += 9) { |
| 2014 | GemmMicrokernelTester() |
| 2015 | .mr(4) |
| 2016 | .nr(8) |
| 2017 | .kr(1) |
| 2018 | .sr(1) |
| 2019 | .m(4) |
| 2020 | .n(8) |
| 2021 | .k(k) |
| 2022 | .b_zero_point(0) |
| 2023 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 2024 | } |
| 2025 | } |
| 2026 | |
| 2027 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_CORTEX_A7, no_zero_point) { |
| 2028 | TEST_REQUIRES_ARM_NEON; |
| 2029 | for (size_t k = 1; k <= 40; k += 9) { |
| 2030 | GemmMicrokernelTester() |
| 2031 | .mr(4) |
| 2032 | .nr(8) |
| 2033 | .kr(1) |
| 2034 | .sr(1) |
| 2035 | .m(4) |
| 2036 | .n(8) |
| 2037 | .k(k) |
| 2038 | .a_zero_point(0) |
| 2039 | .b_zero_point(0) |
| 2040 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_cortex_a7, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
| 2041 | } |
| 2042 | } |
| 2043 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 2044 | |
| 2045 | |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2046 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2047 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_eq_8) { |
| 2048 | TEST_REQUIRES_ARM_NEON; |
| 2049 | GemmMicrokernelTester() |
| 2050 | .mr(2) |
| 2051 | .nr(8) |
| 2052 | .kr(1) |
| 2053 | .sr(1) |
| 2054 | .m(2) |
| 2055 | .n(8) |
| 2056 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2057 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2058 | } |
| 2059 | |
| 2060 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, strided_cn) { |
| 2061 | TEST_REQUIRES_ARM_NEON; |
| 2062 | GemmMicrokernelTester() |
| 2063 | .mr(2) |
| 2064 | .nr(8) |
| 2065 | .kr(1) |
| 2066 | .sr(1) |
| 2067 | .m(2) |
| 2068 | .n(8) |
| 2069 | .k(8) |
| 2070 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2071 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2072 | } |
| 2073 | |
| 2074 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_eq_8_strided_a) { |
| 2075 | TEST_REQUIRES_ARM_NEON; |
| 2076 | GemmMicrokernelTester() |
| 2077 | .mr(2) |
| 2078 | .nr(8) |
| 2079 | .kr(1) |
| 2080 | .sr(1) |
| 2081 | .m(2) |
| 2082 | .n(8) |
| 2083 | .k(8) |
| 2084 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2085 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2086 | } |
| 2087 | |
| 2088 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_eq_8_subtile) { |
| 2089 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2090 | for (uint32_t n = 1; n <= 8; n++) { |
| 2091 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2092 | GemmMicrokernelTester() |
| 2093 | .mr(2) |
| 2094 | .nr(8) |
| 2095 | .kr(1) |
| 2096 | .sr(1) |
| 2097 | .m(m) |
| 2098 | .n(n) |
| 2099 | .k(8) |
| 2100 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2101 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2102 | } |
| 2103 | } |
| 2104 | } |
| 2105 | |
| 2106 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_eq_8_subtile_m) { |
| 2107 | TEST_REQUIRES_ARM_NEON; |
| 2108 | for (uint32_t m = 1; m <= 2; m++) { |
| 2109 | GemmMicrokernelTester() |
| 2110 | .mr(2) |
| 2111 | .nr(8) |
| 2112 | .kr(1) |
| 2113 | .sr(1) |
| 2114 | .m(m) |
| 2115 | .n(8) |
| 2116 | .k(8) |
| 2117 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2118 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2119 | } |
| 2120 | } |
| 2121 | |
| 2122 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_eq_8_subtile_n) { |
| 2123 | TEST_REQUIRES_ARM_NEON; |
| 2124 | for (uint32_t n = 1; n <= 8; n++) { |
| 2125 | GemmMicrokernelTester() |
| 2126 | .mr(2) |
| 2127 | .nr(8) |
| 2128 | .kr(1) |
| 2129 | .sr(1) |
| 2130 | .m(2) |
| 2131 | .n(n) |
| 2132 | .k(8) |
| 2133 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2134 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2135 | } |
| 2136 | } |
| 2137 | |
| 2138 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_lt_8) { |
| 2139 | TEST_REQUIRES_ARM_NEON; |
| 2140 | for (size_t k = 1; k < 8; k++) { |
| 2141 | GemmMicrokernelTester() |
| 2142 | .mr(2) |
| 2143 | .nr(8) |
| 2144 | .kr(1) |
| 2145 | .sr(1) |
| 2146 | .m(2) |
| 2147 | .n(8) |
| 2148 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2149 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2150 | } |
| 2151 | } |
| 2152 | |
| 2153 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_lt_8_strided_a) { |
| 2154 | TEST_REQUIRES_ARM_NEON; |
| 2155 | for (size_t k = 1; k < 8; k++) { |
| 2156 | GemmMicrokernelTester() |
| 2157 | .mr(2) |
| 2158 | .nr(8) |
| 2159 | .kr(1) |
| 2160 | .sr(1) |
| 2161 | .m(2) |
| 2162 | .n(8) |
| 2163 | .k(k) |
| 2164 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2165 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2166 | } |
| 2167 | } |
| 2168 | |
| 2169 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_lt_8_subtile) { |
| 2170 | TEST_REQUIRES_ARM_NEON; |
| 2171 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2172 | for (uint32_t n = 1; n <= 8; n++) { |
| 2173 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2174 | GemmMicrokernelTester() |
| 2175 | .mr(2) |
| 2176 | .nr(8) |
| 2177 | .kr(1) |
| 2178 | .sr(1) |
| 2179 | .m(m) |
| 2180 | .n(n) |
| 2181 | .k(k) |
| 2182 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2183 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2184 | } |
| 2185 | } |
| 2186 | } |
| 2187 | } |
| 2188 | |
| 2189 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_gt_8) { |
| 2190 | TEST_REQUIRES_ARM_NEON; |
| 2191 | for (size_t k = 9; k < 16; k++) { |
| 2192 | GemmMicrokernelTester() |
| 2193 | .mr(2) |
| 2194 | .nr(8) |
| 2195 | .kr(1) |
| 2196 | .sr(1) |
| 2197 | .m(2) |
| 2198 | .n(8) |
| 2199 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2200 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2201 | } |
| 2202 | } |
| 2203 | |
| 2204 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_gt_8_strided_a) { |
| 2205 | TEST_REQUIRES_ARM_NEON; |
| 2206 | for (size_t k = 9; k < 16; k++) { |
| 2207 | GemmMicrokernelTester() |
| 2208 | .mr(2) |
| 2209 | .nr(8) |
| 2210 | .kr(1) |
| 2211 | .sr(1) |
| 2212 | .m(2) |
| 2213 | .n(8) |
| 2214 | .k(k) |
| 2215 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2216 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2217 | } |
| 2218 | } |
| 2219 | |
| 2220 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_gt_8_subtile) { |
| 2221 | TEST_REQUIRES_ARM_NEON; |
| 2222 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2223 | for (uint32_t n = 1; n <= 8; n++) { |
| 2224 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2225 | GemmMicrokernelTester() |
| 2226 | .mr(2) |
| 2227 | .nr(8) |
| 2228 | .kr(1) |
| 2229 | .sr(1) |
| 2230 | .m(m) |
| 2231 | .n(n) |
| 2232 | .k(k) |
| 2233 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2234 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2235 | } |
| 2236 | } |
| 2237 | } |
| 2238 | } |
| 2239 | |
| 2240 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_div_8) { |
| 2241 | TEST_REQUIRES_ARM_NEON; |
| 2242 | for (size_t k = 16; k <= 80; k += 8) { |
| 2243 | GemmMicrokernelTester() |
| 2244 | .mr(2) |
| 2245 | .nr(8) |
| 2246 | .kr(1) |
| 2247 | .sr(1) |
| 2248 | .m(2) |
| 2249 | .n(8) |
| 2250 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2251 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2252 | } |
| 2253 | } |
| 2254 | |
| 2255 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_div_8_strided_a) { |
| 2256 | TEST_REQUIRES_ARM_NEON; |
| 2257 | for (size_t k = 16; k <= 80; k += 8) { |
| 2258 | GemmMicrokernelTester() |
| 2259 | .mr(2) |
| 2260 | .nr(8) |
| 2261 | .kr(1) |
| 2262 | .sr(1) |
| 2263 | .m(2) |
| 2264 | .n(8) |
| 2265 | .k(k) |
| 2266 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2267 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2268 | } |
| 2269 | } |
| 2270 | |
| 2271 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, k_div_8_subtile) { |
| 2272 | TEST_REQUIRES_ARM_NEON; |
| 2273 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2274 | for (uint32_t n = 1; n <= 8; n++) { |
| 2275 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2276 | GemmMicrokernelTester() |
| 2277 | .mr(2) |
| 2278 | .nr(8) |
| 2279 | .kr(1) |
| 2280 | .sr(1) |
| 2281 | .m(m) |
| 2282 | .n(n) |
| 2283 | .k(k) |
| 2284 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2285 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2286 | } |
| 2287 | } |
| 2288 | } |
| 2289 | } |
| 2290 | |
| 2291 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_gt_8) { |
| 2292 | TEST_REQUIRES_ARM_NEON; |
| 2293 | for (uint32_t n = 9; n < 16; n++) { |
| 2294 | for (size_t k = 1; k <= 40; k += 9) { |
| 2295 | GemmMicrokernelTester() |
| 2296 | .mr(2) |
| 2297 | .nr(8) |
| 2298 | .kr(1) |
| 2299 | .sr(1) |
| 2300 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2301 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2302 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2303 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2304 | } |
| 2305 | } |
| 2306 | } |
| 2307 | |
| 2308 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_gt_8_strided_cn) { |
| 2309 | TEST_REQUIRES_ARM_NEON; |
| 2310 | for (uint32_t n = 9; n < 16; n++) { |
| 2311 | for (size_t k = 1; k <= 40; k += 9) { |
| 2312 | GemmMicrokernelTester() |
| 2313 | .mr(2) |
| 2314 | .nr(8) |
| 2315 | .kr(1) |
| 2316 | .sr(1) |
| 2317 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2318 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2319 | .k(k) |
| 2320 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2321 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2322 | } |
| 2323 | } |
| 2324 | } |
| 2325 | |
| 2326 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_gt_8_strided_a) { |
| 2327 | TEST_REQUIRES_ARM_NEON; |
| 2328 | for (uint32_t n = 9; n < 16; n++) { |
| 2329 | for (size_t k = 1; k <= 40; k += 9) { |
| 2330 | GemmMicrokernelTester() |
| 2331 | .mr(2) |
| 2332 | .nr(8) |
| 2333 | .kr(1) |
| 2334 | .sr(1) |
| 2335 | .m(2) |
| 2336 | .n(n) |
| 2337 | .k(k) |
| 2338 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2339 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2340 | } |
| 2341 | } |
| 2342 | } |
| 2343 | |
| 2344 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_gt_8_subtile) { |
| 2345 | TEST_REQUIRES_ARM_NEON; |
| 2346 | for (uint32_t n = 9; n < 16; n++) { |
| 2347 | for (size_t k = 1; k <= 40; k += 9) { |
| 2348 | for (uint32_t m = 1; m <= 2; m++) { |
| 2349 | GemmMicrokernelTester() |
| 2350 | .mr(2) |
| 2351 | .nr(8) |
| 2352 | .kr(1) |
| 2353 | .sr(1) |
| 2354 | .m(m) |
| 2355 | .n(n) |
| 2356 | .k(k) |
| 2357 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2358 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2359 | } |
| 2360 | } |
| 2361 | } |
| 2362 | } |
| 2363 | |
| 2364 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_div_8) { |
| 2365 | TEST_REQUIRES_ARM_NEON; |
| 2366 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2367 | for (size_t k = 1; k <= 40; k += 9) { |
| 2368 | GemmMicrokernelTester() |
| 2369 | .mr(2) |
| 2370 | .nr(8) |
| 2371 | .kr(1) |
| 2372 | .sr(1) |
| 2373 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2374 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2375 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2376 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2377 | } |
| 2378 | } |
| 2379 | } |
| 2380 | |
| 2381 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_div_8_strided_cn) { |
| 2382 | TEST_REQUIRES_ARM_NEON; |
| 2383 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2384 | for (size_t k = 1; k <= 40; k += 9) { |
| 2385 | GemmMicrokernelTester() |
| 2386 | .mr(2) |
| 2387 | .nr(8) |
| 2388 | .kr(1) |
| 2389 | .sr(1) |
| 2390 | .m(2) |
| 2391 | .n(n) |
| 2392 | .k(k) |
| 2393 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2394 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2395 | } |
| 2396 | } |
| 2397 | } |
| 2398 | |
| 2399 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_div_8_strided_a) { |
| 2400 | TEST_REQUIRES_ARM_NEON; |
| 2401 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2402 | for (size_t k = 1; k <= 40; k += 9) { |
| 2403 | GemmMicrokernelTester() |
| 2404 | .mr(2) |
| 2405 | .nr(8) |
| 2406 | .kr(1) |
| 2407 | .sr(1) |
| 2408 | .m(2) |
| 2409 | .n(n) |
| 2410 | .k(k) |
| 2411 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2412 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2413 | } |
| 2414 | } |
| 2415 | } |
| 2416 | |
| 2417 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, n_div_8_subtile) { |
| 2418 | TEST_REQUIRES_ARM_NEON; |
| 2419 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2420 | for (size_t k = 1; k <= 40; k += 9) { |
| 2421 | for (uint32_t m = 1; m <= 2; m++) { |
| 2422 | GemmMicrokernelTester() |
| 2423 | .mr(2) |
| 2424 | .nr(8) |
| 2425 | .kr(1) |
| 2426 | .sr(1) |
| 2427 | .m(m) |
| 2428 | .n(n) |
| 2429 | .k(k) |
| 2430 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2431 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2432 | } |
| 2433 | } |
| 2434 | } |
| 2435 | } |
| 2436 | |
| 2437 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, strided_cm_subtile) { |
| 2438 | TEST_REQUIRES_ARM_NEON; |
| 2439 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2440 | for (uint32_t n = 1; n <= 8; n++) { |
| 2441 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2442 | GemmMicrokernelTester() |
| 2443 | .mr(2) |
| 2444 | .nr(8) |
| 2445 | .kr(1) |
| 2446 | .sr(1) |
| 2447 | .m(m) |
| 2448 | .n(n) |
| 2449 | .k(k) |
| 2450 | .cm_stride(11) |
| 2451 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2452 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2453 | } |
| 2454 | } |
| 2455 | } |
| 2456 | } |
| 2457 | |
| 2458 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, qmin) { |
| 2459 | TEST_REQUIRES_ARM_NEON; |
| 2460 | GemmMicrokernelTester() |
| 2461 | .mr(2) |
| 2462 | .nr(8) |
| 2463 | .kr(1) |
| 2464 | .sr(1) |
| 2465 | .m(2) |
| 2466 | .n(8) |
| 2467 | .k(8) |
| 2468 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2469 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2470 | } |
| 2471 | |
| 2472 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, qmax) { |
| 2473 | TEST_REQUIRES_ARM_NEON; |
| 2474 | GemmMicrokernelTester() |
| 2475 | .mr(2) |
| 2476 | .nr(8) |
| 2477 | .kr(1) |
| 2478 | .sr(1) |
| 2479 | .m(2) |
| 2480 | .n(8) |
| 2481 | .k(8) |
| 2482 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2483 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2484 | } |
| 2485 | |
| 2486 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, strided_cm) { |
| 2487 | TEST_REQUIRES_ARM_NEON; |
| 2488 | GemmMicrokernelTester() |
| 2489 | .mr(2) |
| 2490 | .nr(8) |
| 2491 | .kr(1) |
| 2492 | .sr(1) |
| 2493 | .m(2) |
| 2494 | .n(8) |
| 2495 | .k(8) |
| 2496 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2497 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2498 | } |
| 2499 | |
| 2500 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, no_a_zero_point) { |
| 2501 | TEST_REQUIRES_ARM_NEON; |
| 2502 | for (size_t k = 1; k <= 40; k += 9) { |
| 2503 | GemmMicrokernelTester() |
| 2504 | .mr(2) |
| 2505 | .nr(8) |
| 2506 | .kr(1) |
| 2507 | .sr(1) |
| 2508 | .m(2) |
| 2509 | .n(8) |
| 2510 | .k(k) |
| 2511 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2512 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2513 | } |
| 2514 | } |
| 2515 | |
| 2516 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, no_b_zero_point) { |
| 2517 | TEST_REQUIRES_ARM_NEON; |
| 2518 | for (size_t k = 1; k <= 40; k += 9) { |
| 2519 | GemmMicrokernelTester() |
| 2520 | .mr(2) |
| 2521 | .nr(8) |
| 2522 | .kr(1) |
| 2523 | .sr(1) |
| 2524 | .m(2) |
| 2525 | .n(8) |
| 2526 | .k(k) |
| 2527 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2528 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2529 | } |
| 2530 | } |
| 2531 | |
| 2532 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8__NEON_MLAL_LANE, no_zero_point) { |
| 2533 | TEST_REQUIRES_ARM_NEON; |
| 2534 | for (size_t k = 1; k <= 40; k += 9) { |
| 2535 | GemmMicrokernelTester() |
| 2536 | .mr(2) |
| 2537 | .nr(8) |
| 2538 | .kr(1) |
| 2539 | .sr(1) |
| 2540 | .m(2) |
| 2541 | .n(8) |
| 2542 | .k(k) |
| 2543 | .a_zero_point(0) |
| 2544 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2545 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2546 | } |
| 2547 | } |
| 2548 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 2549 | |
| 2550 | |
| 2551 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 2552 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_eq_8) { |
| 2553 | TEST_REQUIRES_ARM_NEON; |
| 2554 | GemmMicrokernelTester() |
| 2555 | .mr(3) |
| 2556 | .nr(8) |
| 2557 | .kr(1) |
| 2558 | .sr(1) |
| 2559 | .m(3) |
| 2560 | .n(8) |
| 2561 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2562 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2563 | } |
| 2564 | |
| 2565 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, strided_cn) { |
| 2566 | TEST_REQUIRES_ARM_NEON; |
| 2567 | GemmMicrokernelTester() |
| 2568 | .mr(3) |
| 2569 | .nr(8) |
| 2570 | .kr(1) |
| 2571 | .sr(1) |
| 2572 | .m(3) |
| 2573 | .n(8) |
| 2574 | .k(8) |
| 2575 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2576 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2577 | } |
| 2578 | |
| 2579 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_eq_8_strided_a) { |
| 2580 | TEST_REQUIRES_ARM_NEON; |
| 2581 | GemmMicrokernelTester() |
| 2582 | .mr(3) |
| 2583 | .nr(8) |
| 2584 | .kr(1) |
| 2585 | .sr(1) |
| 2586 | .m(3) |
| 2587 | .n(8) |
| 2588 | .k(8) |
| 2589 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2590 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2591 | } |
| 2592 | |
| 2593 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_eq_8_subtile) { |
| 2594 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2595 | for (uint32_t n = 1; n <= 8; n++) { |
| 2596 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2597 | GemmMicrokernelTester() |
| 2598 | .mr(3) |
| 2599 | .nr(8) |
| 2600 | .kr(1) |
| 2601 | .sr(1) |
| 2602 | .m(m) |
| 2603 | .n(n) |
| 2604 | .k(8) |
| 2605 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2606 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2607 | } |
| 2608 | } |
| 2609 | } |
| 2610 | |
| 2611 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_eq_8_subtile_m) { |
| 2612 | TEST_REQUIRES_ARM_NEON; |
| 2613 | for (uint32_t m = 1; m <= 3; m++) { |
| 2614 | GemmMicrokernelTester() |
| 2615 | .mr(3) |
| 2616 | .nr(8) |
| 2617 | .kr(1) |
| 2618 | .sr(1) |
| 2619 | .m(m) |
| 2620 | .n(8) |
| 2621 | .k(8) |
| 2622 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2623 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2624 | } |
| 2625 | } |
| 2626 | |
| 2627 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_eq_8_subtile_n) { |
| 2628 | TEST_REQUIRES_ARM_NEON; |
| 2629 | for (uint32_t n = 1; n <= 8; n++) { |
| 2630 | GemmMicrokernelTester() |
| 2631 | .mr(3) |
| 2632 | .nr(8) |
| 2633 | .kr(1) |
| 2634 | .sr(1) |
| 2635 | .m(3) |
| 2636 | .n(n) |
| 2637 | .k(8) |
| 2638 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2639 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2640 | } |
| 2641 | } |
| 2642 | |
| 2643 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_lt_8) { |
| 2644 | TEST_REQUIRES_ARM_NEON; |
| 2645 | for (size_t k = 1; k < 8; k++) { |
| 2646 | GemmMicrokernelTester() |
| 2647 | .mr(3) |
| 2648 | .nr(8) |
| 2649 | .kr(1) |
| 2650 | .sr(1) |
| 2651 | .m(3) |
| 2652 | .n(8) |
| 2653 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2654 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2655 | } |
| 2656 | } |
| 2657 | |
| 2658 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_lt_8_strided_a) { |
| 2659 | TEST_REQUIRES_ARM_NEON; |
| 2660 | for (size_t k = 1; k < 8; k++) { |
| 2661 | GemmMicrokernelTester() |
| 2662 | .mr(3) |
| 2663 | .nr(8) |
| 2664 | .kr(1) |
| 2665 | .sr(1) |
| 2666 | .m(3) |
| 2667 | .n(8) |
| 2668 | .k(k) |
| 2669 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2670 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2671 | } |
| 2672 | } |
| 2673 | |
| 2674 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_lt_8_subtile) { |
| 2675 | TEST_REQUIRES_ARM_NEON; |
| 2676 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2677 | for (uint32_t n = 1; n <= 8; n++) { |
| 2678 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2679 | GemmMicrokernelTester() |
| 2680 | .mr(3) |
| 2681 | .nr(8) |
| 2682 | .kr(1) |
| 2683 | .sr(1) |
| 2684 | .m(m) |
| 2685 | .n(n) |
| 2686 | .k(k) |
| 2687 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2688 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2689 | } |
| 2690 | } |
| 2691 | } |
| 2692 | } |
| 2693 | |
| 2694 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_gt_8) { |
| 2695 | TEST_REQUIRES_ARM_NEON; |
| 2696 | for (size_t k = 9; k < 16; k++) { |
| 2697 | GemmMicrokernelTester() |
| 2698 | .mr(3) |
| 2699 | .nr(8) |
| 2700 | .kr(1) |
| 2701 | .sr(1) |
| 2702 | .m(3) |
| 2703 | .n(8) |
| 2704 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2705 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2706 | } |
| 2707 | } |
| 2708 | |
| 2709 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_gt_8_strided_a) { |
| 2710 | TEST_REQUIRES_ARM_NEON; |
| 2711 | for (size_t k = 9; k < 16; k++) { |
| 2712 | GemmMicrokernelTester() |
| 2713 | .mr(3) |
| 2714 | .nr(8) |
| 2715 | .kr(1) |
| 2716 | .sr(1) |
| 2717 | .m(3) |
| 2718 | .n(8) |
| 2719 | .k(k) |
| 2720 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2721 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2722 | } |
| 2723 | } |
| 2724 | |
| 2725 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_gt_8_subtile) { |
| 2726 | TEST_REQUIRES_ARM_NEON; |
| 2727 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2728 | for (uint32_t n = 1; n <= 8; n++) { |
| 2729 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2730 | GemmMicrokernelTester() |
| 2731 | .mr(3) |
| 2732 | .nr(8) |
| 2733 | .kr(1) |
| 2734 | .sr(1) |
| 2735 | .m(m) |
| 2736 | .n(n) |
| 2737 | .k(k) |
| 2738 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2739 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2740 | } |
| 2741 | } |
| 2742 | } |
| 2743 | } |
| 2744 | |
| 2745 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_div_8) { |
| 2746 | TEST_REQUIRES_ARM_NEON; |
| 2747 | for (size_t k = 16; k <= 80; k += 8) { |
| 2748 | GemmMicrokernelTester() |
| 2749 | .mr(3) |
| 2750 | .nr(8) |
| 2751 | .kr(1) |
| 2752 | .sr(1) |
| 2753 | .m(3) |
| 2754 | .n(8) |
| 2755 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2756 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2757 | } |
| 2758 | } |
| 2759 | |
| 2760 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_div_8_strided_a) { |
| 2761 | TEST_REQUIRES_ARM_NEON; |
| 2762 | for (size_t k = 16; k <= 80; k += 8) { |
| 2763 | GemmMicrokernelTester() |
| 2764 | .mr(3) |
| 2765 | .nr(8) |
| 2766 | .kr(1) |
| 2767 | .sr(1) |
| 2768 | .m(3) |
| 2769 | .n(8) |
| 2770 | .k(k) |
| 2771 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2772 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2773 | } |
| 2774 | } |
| 2775 | |
| 2776 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, k_div_8_subtile) { |
| 2777 | TEST_REQUIRES_ARM_NEON; |
| 2778 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2779 | for (uint32_t n = 1; n <= 8; n++) { |
| 2780 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2781 | GemmMicrokernelTester() |
| 2782 | .mr(3) |
| 2783 | .nr(8) |
| 2784 | .kr(1) |
| 2785 | .sr(1) |
| 2786 | .m(m) |
| 2787 | .n(n) |
| 2788 | .k(k) |
| 2789 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2790 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2791 | } |
| 2792 | } |
| 2793 | } |
| 2794 | } |
| 2795 | |
| 2796 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_gt_8) { |
| 2797 | TEST_REQUIRES_ARM_NEON; |
| 2798 | for (uint32_t n = 9; n < 16; n++) { |
| 2799 | for (size_t k = 1; k <= 40; k += 9) { |
| 2800 | GemmMicrokernelTester() |
| 2801 | .mr(3) |
| 2802 | .nr(8) |
| 2803 | .kr(1) |
| 2804 | .sr(1) |
| 2805 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2806 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2807 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2808 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2809 | } |
| 2810 | } |
| 2811 | } |
| 2812 | |
| 2813 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_gt_8_strided_cn) { |
| 2814 | TEST_REQUIRES_ARM_NEON; |
| 2815 | for (uint32_t n = 9; n < 16; n++) { |
| 2816 | for (size_t k = 1; k <= 40; k += 9) { |
| 2817 | GemmMicrokernelTester() |
| 2818 | .mr(3) |
| 2819 | .nr(8) |
| 2820 | .kr(1) |
| 2821 | .sr(1) |
| 2822 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2823 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2824 | .k(k) |
| 2825 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2826 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2827 | } |
| 2828 | } |
| 2829 | } |
| 2830 | |
| 2831 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_gt_8_strided_a) { |
| 2832 | TEST_REQUIRES_ARM_NEON; |
| 2833 | for (uint32_t n = 9; n < 16; n++) { |
| 2834 | for (size_t k = 1; k <= 40; k += 9) { |
| 2835 | GemmMicrokernelTester() |
| 2836 | .mr(3) |
| 2837 | .nr(8) |
| 2838 | .kr(1) |
| 2839 | .sr(1) |
| 2840 | .m(3) |
| 2841 | .n(n) |
| 2842 | .k(k) |
| 2843 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2844 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2845 | } |
| 2846 | } |
| 2847 | } |
| 2848 | |
| 2849 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_gt_8_subtile) { |
| 2850 | TEST_REQUIRES_ARM_NEON; |
| 2851 | for (uint32_t n = 9; n < 16; n++) { |
| 2852 | for (size_t k = 1; k <= 40; k += 9) { |
| 2853 | for (uint32_t m = 1; m <= 3; m++) { |
| 2854 | GemmMicrokernelTester() |
| 2855 | .mr(3) |
| 2856 | .nr(8) |
| 2857 | .kr(1) |
| 2858 | .sr(1) |
| 2859 | .m(m) |
| 2860 | .n(n) |
| 2861 | .k(k) |
| 2862 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2863 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2864 | } |
| 2865 | } |
| 2866 | } |
| 2867 | } |
| 2868 | |
| 2869 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_div_8) { |
| 2870 | TEST_REQUIRES_ARM_NEON; |
| 2871 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2872 | for (size_t k = 1; k <= 40; k += 9) { |
| 2873 | GemmMicrokernelTester() |
| 2874 | .mr(3) |
| 2875 | .nr(8) |
| 2876 | .kr(1) |
| 2877 | .sr(1) |
| 2878 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2879 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2880 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2881 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2882 | } |
| 2883 | } |
| 2884 | } |
| 2885 | |
| 2886 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_div_8_strided_cn) { |
| 2887 | TEST_REQUIRES_ARM_NEON; |
| 2888 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2889 | for (size_t k = 1; k <= 40; k += 9) { |
| 2890 | GemmMicrokernelTester() |
| 2891 | .mr(3) |
| 2892 | .nr(8) |
| 2893 | .kr(1) |
| 2894 | .sr(1) |
| 2895 | .m(3) |
| 2896 | .n(n) |
| 2897 | .k(k) |
| 2898 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2899 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2900 | } |
| 2901 | } |
| 2902 | } |
| 2903 | |
| 2904 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_div_8_strided_a) { |
| 2905 | TEST_REQUIRES_ARM_NEON; |
| 2906 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2907 | for (size_t k = 1; k <= 40; k += 9) { |
| 2908 | GemmMicrokernelTester() |
| 2909 | .mr(3) |
| 2910 | .nr(8) |
| 2911 | .kr(1) |
| 2912 | .sr(1) |
| 2913 | .m(3) |
| 2914 | .n(n) |
| 2915 | .k(k) |
| 2916 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2917 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2918 | } |
| 2919 | } |
| 2920 | } |
| 2921 | |
| 2922 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, n_div_8_subtile) { |
| 2923 | TEST_REQUIRES_ARM_NEON; |
| 2924 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 2925 | for (size_t k = 1; k <= 40; k += 9) { |
| 2926 | for (uint32_t m = 1; m <= 3; m++) { |
| 2927 | GemmMicrokernelTester() |
| 2928 | .mr(3) |
| 2929 | .nr(8) |
| 2930 | .kr(1) |
| 2931 | .sr(1) |
| 2932 | .m(m) |
| 2933 | .n(n) |
| 2934 | .k(k) |
| 2935 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2936 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2937 | } |
| 2938 | } |
| 2939 | } |
| 2940 | } |
| 2941 | |
| 2942 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, strided_cm_subtile) { |
| 2943 | TEST_REQUIRES_ARM_NEON; |
| 2944 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2945 | for (uint32_t n = 1; n <= 8; n++) { |
| 2946 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2947 | GemmMicrokernelTester() |
| 2948 | .mr(3) |
| 2949 | .nr(8) |
| 2950 | .kr(1) |
| 2951 | .sr(1) |
| 2952 | .m(m) |
| 2953 | .n(n) |
| 2954 | .k(k) |
| 2955 | .cm_stride(11) |
| 2956 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2957 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2958 | } |
| 2959 | } |
| 2960 | } |
| 2961 | } |
| 2962 | |
| 2963 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, qmin) { |
| 2964 | TEST_REQUIRES_ARM_NEON; |
| 2965 | GemmMicrokernelTester() |
| 2966 | .mr(3) |
| 2967 | .nr(8) |
| 2968 | .kr(1) |
| 2969 | .sr(1) |
| 2970 | .m(3) |
| 2971 | .n(8) |
| 2972 | .k(8) |
| 2973 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2974 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2975 | } |
| 2976 | |
| 2977 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, qmax) { |
| 2978 | TEST_REQUIRES_ARM_NEON; |
| 2979 | GemmMicrokernelTester() |
| 2980 | .mr(3) |
| 2981 | .nr(8) |
| 2982 | .kr(1) |
| 2983 | .sr(1) |
| 2984 | .m(3) |
| 2985 | .n(8) |
| 2986 | .k(8) |
| 2987 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2988 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 2989 | } |
| 2990 | |
| 2991 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, strided_cm) { |
| 2992 | TEST_REQUIRES_ARM_NEON; |
| 2993 | GemmMicrokernelTester() |
| 2994 | .mr(3) |
| 2995 | .nr(8) |
| 2996 | .kr(1) |
| 2997 | .sr(1) |
| 2998 | .m(3) |
| 2999 | .n(8) |
| 3000 | .k(8) |
| 3001 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3002 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3003 | } |
| 3004 | |
| 3005 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, no_a_zero_point) { |
| 3006 | TEST_REQUIRES_ARM_NEON; |
| 3007 | for (size_t k = 1; k <= 40; k += 9) { |
| 3008 | GemmMicrokernelTester() |
| 3009 | .mr(3) |
| 3010 | .nr(8) |
| 3011 | .kr(1) |
| 3012 | .sr(1) |
| 3013 | .m(3) |
| 3014 | .n(8) |
| 3015 | .k(k) |
| 3016 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3017 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3018 | } |
| 3019 | } |
| 3020 | |
| 3021 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, no_b_zero_point) { |
| 3022 | TEST_REQUIRES_ARM_NEON; |
| 3023 | for (size_t k = 1; k <= 40; k += 9) { |
| 3024 | GemmMicrokernelTester() |
| 3025 | .mr(3) |
| 3026 | .nr(8) |
| 3027 | .kr(1) |
| 3028 | .sr(1) |
| 3029 | .m(3) |
| 3030 | .n(8) |
| 3031 | .k(k) |
| 3032 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3033 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3034 | } |
| 3035 | } |
| 3036 | |
| 3037 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8__NEON_MLAL_LANE, no_zero_point) { |
| 3038 | TEST_REQUIRES_ARM_NEON; |
| 3039 | for (size_t k = 1; k <= 40; k += 9) { |
| 3040 | GemmMicrokernelTester() |
| 3041 | .mr(3) |
| 3042 | .nr(8) |
| 3043 | .kr(1) |
| 3044 | .sr(1) |
| 3045 | .m(3) |
| 3046 | .n(8) |
| 3047 | .k(k) |
| 3048 | .a_zero_point(0) |
| 3049 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3050 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3051 | } |
| 3052 | } |
| 3053 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 3054 | |
| 3055 | |
| 3056 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3057 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8) { |
| 3058 | TEST_REQUIRES_ARM_NEON; |
| 3059 | GemmMicrokernelTester() |
| 3060 | .mr(6) |
| 3061 | .nr(8) |
| 3062 | .kr(1) |
| 3063 | .sr(1) |
| 3064 | .m(6) |
| 3065 | .n(8) |
| 3066 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3067 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3068 | } |
| 3069 | |
| 3070 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, strided_cn) { |
| 3071 | TEST_REQUIRES_ARM_NEON; |
| 3072 | GemmMicrokernelTester() |
| 3073 | .mr(6) |
| 3074 | .nr(8) |
| 3075 | .kr(1) |
| 3076 | .sr(1) |
| 3077 | .m(6) |
| 3078 | .n(8) |
| 3079 | .k(8) |
| 3080 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3081 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3082 | } |
| 3083 | |
| 3084 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8_strided_a) { |
| 3085 | TEST_REQUIRES_ARM_NEON; |
| 3086 | GemmMicrokernelTester() |
| 3087 | .mr(6) |
| 3088 | .nr(8) |
| 3089 | .kr(1) |
| 3090 | .sr(1) |
| 3091 | .m(6) |
| 3092 | .n(8) |
| 3093 | .k(8) |
| 3094 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3095 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3096 | } |
| 3097 | |
| 3098 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8_subtile) { |
| 3099 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3100 | for (uint32_t n = 1; n <= 8; n++) { |
| 3101 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3102 | GemmMicrokernelTester() |
| 3103 | .mr(6) |
| 3104 | .nr(8) |
| 3105 | .kr(1) |
| 3106 | .sr(1) |
| 3107 | .m(m) |
| 3108 | .n(n) |
| 3109 | .k(8) |
| 3110 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3111 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3112 | } |
| 3113 | } |
| 3114 | } |
| 3115 | |
| 3116 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8_subtile_m) { |
| 3117 | TEST_REQUIRES_ARM_NEON; |
| 3118 | for (uint32_t m = 1; m <= 6; m++) { |
| 3119 | GemmMicrokernelTester() |
| 3120 | .mr(6) |
| 3121 | .nr(8) |
| 3122 | .kr(1) |
| 3123 | .sr(1) |
| 3124 | .m(m) |
| 3125 | .n(8) |
| 3126 | .k(8) |
| 3127 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3128 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3129 | } |
| 3130 | } |
| 3131 | |
| 3132 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8_subtile_n) { |
| 3133 | TEST_REQUIRES_ARM_NEON; |
| 3134 | for (uint32_t n = 1; n <= 8; n++) { |
| 3135 | GemmMicrokernelTester() |
| 3136 | .mr(6) |
| 3137 | .nr(8) |
| 3138 | .kr(1) |
| 3139 | .sr(1) |
| 3140 | .m(6) |
| 3141 | .n(n) |
| 3142 | .k(8) |
| 3143 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3144 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3145 | } |
| 3146 | } |
| 3147 | |
| 3148 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_lt_8) { |
| 3149 | TEST_REQUIRES_ARM_NEON; |
| 3150 | for (size_t k = 1; k < 8; k++) { |
| 3151 | GemmMicrokernelTester() |
| 3152 | .mr(6) |
| 3153 | .nr(8) |
| 3154 | .kr(1) |
| 3155 | .sr(1) |
| 3156 | .m(6) |
| 3157 | .n(8) |
| 3158 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3159 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3160 | } |
| 3161 | } |
| 3162 | |
| 3163 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_lt_8_strided_a) { |
| 3164 | TEST_REQUIRES_ARM_NEON; |
| 3165 | for (size_t k = 1; k < 8; k++) { |
| 3166 | GemmMicrokernelTester() |
| 3167 | .mr(6) |
| 3168 | .nr(8) |
| 3169 | .kr(1) |
| 3170 | .sr(1) |
| 3171 | .m(6) |
| 3172 | .n(8) |
| 3173 | .k(k) |
| 3174 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3175 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3176 | } |
| 3177 | } |
| 3178 | |
| 3179 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_lt_8_subtile) { |
| 3180 | TEST_REQUIRES_ARM_NEON; |
| 3181 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3182 | for (uint32_t n = 1; n <= 8; n++) { |
| 3183 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3184 | GemmMicrokernelTester() |
| 3185 | .mr(6) |
| 3186 | .nr(8) |
| 3187 | .kr(1) |
| 3188 | .sr(1) |
| 3189 | .m(m) |
| 3190 | .n(n) |
| 3191 | .k(k) |
| 3192 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3193 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3194 | } |
| 3195 | } |
| 3196 | } |
| 3197 | } |
| 3198 | |
| 3199 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_gt_8) { |
| 3200 | TEST_REQUIRES_ARM_NEON; |
| 3201 | for (size_t k = 9; k < 16; k++) { |
| 3202 | GemmMicrokernelTester() |
| 3203 | .mr(6) |
| 3204 | .nr(8) |
| 3205 | .kr(1) |
| 3206 | .sr(1) |
| 3207 | .m(6) |
| 3208 | .n(8) |
| 3209 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3210 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3211 | } |
| 3212 | } |
| 3213 | |
| 3214 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_gt_8_strided_a) { |
| 3215 | TEST_REQUIRES_ARM_NEON; |
| 3216 | for (size_t k = 9; k < 16; k++) { |
| 3217 | GemmMicrokernelTester() |
| 3218 | .mr(6) |
| 3219 | .nr(8) |
| 3220 | .kr(1) |
| 3221 | .sr(1) |
| 3222 | .m(6) |
| 3223 | .n(8) |
| 3224 | .k(k) |
| 3225 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3226 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3227 | } |
| 3228 | } |
| 3229 | |
| 3230 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_gt_8_subtile) { |
| 3231 | TEST_REQUIRES_ARM_NEON; |
| 3232 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3233 | for (uint32_t n = 1; n <= 8; n++) { |
| 3234 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3235 | GemmMicrokernelTester() |
| 3236 | .mr(6) |
| 3237 | .nr(8) |
| 3238 | .kr(1) |
| 3239 | .sr(1) |
| 3240 | .m(m) |
| 3241 | .n(n) |
| 3242 | .k(k) |
| 3243 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3244 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3245 | } |
| 3246 | } |
| 3247 | } |
| 3248 | } |
| 3249 | |
| 3250 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_div_8) { |
| 3251 | TEST_REQUIRES_ARM_NEON; |
| 3252 | for (size_t k = 16; k <= 80; k += 8) { |
| 3253 | GemmMicrokernelTester() |
| 3254 | .mr(6) |
| 3255 | .nr(8) |
| 3256 | .kr(1) |
| 3257 | .sr(1) |
| 3258 | .m(6) |
| 3259 | .n(8) |
| 3260 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3261 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3262 | } |
| 3263 | } |
| 3264 | |
| 3265 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_div_8_strided_a) { |
| 3266 | TEST_REQUIRES_ARM_NEON; |
| 3267 | for (size_t k = 16; k <= 80; k += 8) { |
| 3268 | GemmMicrokernelTester() |
| 3269 | .mr(6) |
| 3270 | .nr(8) |
| 3271 | .kr(1) |
| 3272 | .sr(1) |
| 3273 | .m(6) |
| 3274 | .n(8) |
| 3275 | .k(k) |
| 3276 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3277 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3278 | } |
| 3279 | } |
| 3280 | |
| 3281 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_div_8_subtile) { |
| 3282 | TEST_REQUIRES_ARM_NEON; |
| 3283 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3284 | for (uint32_t n = 1; n <= 8; n++) { |
| 3285 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3286 | GemmMicrokernelTester() |
| 3287 | .mr(6) |
| 3288 | .nr(8) |
| 3289 | .kr(1) |
| 3290 | .sr(1) |
| 3291 | .m(m) |
| 3292 | .n(n) |
| 3293 | .k(k) |
| 3294 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3295 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3296 | } |
| 3297 | } |
| 3298 | } |
| 3299 | } |
| 3300 | |
| 3301 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8) { |
| 3302 | TEST_REQUIRES_ARM_NEON; |
| 3303 | for (uint32_t n = 9; n < 16; n++) { |
| 3304 | for (size_t k = 1; k <= 40; k += 9) { |
| 3305 | GemmMicrokernelTester() |
| 3306 | .mr(6) |
| 3307 | .nr(8) |
| 3308 | .kr(1) |
| 3309 | .sr(1) |
| 3310 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3311 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3312 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3313 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3314 | } |
| 3315 | } |
| 3316 | } |
| 3317 | |
| 3318 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8_strided_cn) { |
| 3319 | TEST_REQUIRES_ARM_NEON; |
| 3320 | for (uint32_t n = 9; n < 16; n++) { |
| 3321 | for (size_t k = 1; k <= 40; k += 9) { |
| 3322 | GemmMicrokernelTester() |
| 3323 | .mr(6) |
| 3324 | .nr(8) |
| 3325 | .kr(1) |
| 3326 | .sr(1) |
| 3327 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3328 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3329 | .k(k) |
| 3330 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3331 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3332 | } |
| 3333 | } |
| 3334 | } |
| 3335 | |
| 3336 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8_strided_a) { |
| 3337 | TEST_REQUIRES_ARM_NEON; |
| 3338 | for (uint32_t n = 9; n < 16; n++) { |
| 3339 | for (size_t k = 1; k <= 40; k += 9) { |
| 3340 | GemmMicrokernelTester() |
| 3341 | .mr(6) |
| 3342 | .nr(8) |
| 3343 | .kr(1) |
| 3344 | .sr(1) |
| 3345 | .m(6) |
| 3346 | .n(n) |
| 3347 | .k(k) |
| 3348 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3349 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3350 | } |
| 3351 | } |
| 3352 | } |
| 3353 | |
| 3354 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8_subtile) { |
| 3355 | TEST_REQUIRES_ARM_NEON; |
| 3356 | for (uint32_t n = 9; n < 16; n++) { |
| 3357 | for (size_t k = 1; k <= 40; k += 9) { |
| 3358 | for (uint32_t m = 1; m <= 6; m++) { |
| 3359 | GemmMicrokernelTester() |
| 3360 | .mr(6) |
| 3361 | .nr(8) |
| 3362 | .kr(1) |
| 3363 | .sr(1) |
| 3364 | .m(m) |
| 3365 | .n(n) |
| 3366 | .k(k) |
| 3367 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3368 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3369 | } |
| 3370 | } |
| 3371 | } |
| 3372 | } |
| 3373 | |
| 3374 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8) { |
| 3375 | TEST_REQUIRES_ARM_NEON; |
| 3376 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3377 | for (size_t k = 1; k <= 40; k += 9) { |
| 3378 | GemmMicrokernelTester() |
| 3379 | .mr(6) |
| 3380 | .nr(8) |
| 3381 | .kr(1) |
| 3382 | .sr(1) |
| 3383 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3384 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3385 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3386 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3387 | } |
| 3388 | } |
| 3389 | } |
| 3390 | |
| 3391 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8_strided_cn) { |
| 3392 | TEST_REQUIRES_ARM_NEON; |
| 3393 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3394 | for (size_t k = 1; k <= 40; k += 9) { |
| 3395 | GemmMicrokernelTester() |
| 3396 | .mr(6) |
| 3397 | .nr(8) |
| 3398 | .kr(1) |
| 3399 | .sr(1) |
| 3400 | .m(6) |
| 3401 | .n(n) |
| 3402 | .k(k) |
| 3403 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3404 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3405 | } |
| 3406 | } |
| 3407 | } |
| 3408 | |
| 3409 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8_strided_a) { |
| 3410 | TEST_REQUIRES_ARM_NEON; |
| 3411 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3412 | for (size_t k = 1; k <= 40; k += 9) { |
| 3413 | GemmMicrokernelTester() |
| 3414 | .mr(6) |
| 3415 | .nr(8) |
| 3416 | .kr(1) |
| 3417 | .sr(1) |
| 3418 | .m(6) |
| 3419 | .n(n) |
| 3420 | .k(k) |
| 3421 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3422 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3423 | } |
| 3424 | } |
| 3425 | } |
| 3426 | |
| 3427 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8_subtile) { |
| 3428 | TEST_REQUIRES_ARM_NEON; |
| 3429 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3430 | for (size_t k = 1; k <= 40; k += 9) { |
| 3431 | for (uint32_t m = 1; m <= 6; m++) { |
| 3432 | GemmMicrokernelTester() |
| 3433 | .mr(6) |
| 3434 | .nr(8) |
| 3435 | .kr(1) |
| 3436 | .sr(1) |
| 3437 | .m(m) |
| 3438 | .n(n) |
| 3439 | .k(k) |
| 3440 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3441 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3442 | } |
| 3443 | } |
| 3444 | } |
| 3445 | } |
| 3446 | |
| 3447 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, strided_cm_subtile) { |
| 3448 | TEST_REQUIRES_ARM_NEON; |
| 3449 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3450 | for (uint32_t n = 1; n <= 8; n++) { |
| 3451 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3452 | GemmMicrokernelTester() |
| 3453 | .mr(6) |
| 3454 | .nr(8) |
| 3455 | .kr(1) |
| 3456 | .sr(1) |
| 3457 | .m(m) |
| 3458 | .n(n) |
| 3459 | .k(k) |
| 3460 | .cm_stride(11) |
| 3461 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3462 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3463 | } |
| 3464 | } |
| 3465 | } |
| 3466 | } |
| 3467 | |
| 3468 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, qmin) { |
| 3469 | TEST_REQUIRES_ARM_NEON; |
| 3470 | GemmMicrokernelTester() |
| 3471 | .mr(6) |
| 3472 | .nr(8) |
| 3473 | .kr(1) |
| 3474 | .sr(1) |
| 3475 | .m(6) |
| 3476 | .n(8) |
| 3477 | .k(8) |
| 3478 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3479 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3480 | } |
| 3481 | |
| 3482 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, qmax) { |
| 3483 | TEST_REQUIRES_ARM_NEON; |
| 3484 | GemmMicrokernelTester() |
| 3485 | .mr(6) |
| 3486 | .nr(8) |
| 3487 | .kr(1) |
| 3488 | .sr(1) |
| 3489 | .m(6) |
| 3490 | .n(8) |
| 3491 | .k(8) |
| 3492 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3493 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3494 | } |
| 3495 | |
| 3496 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, strided_cm) { |
| 3497 | TEST_REQUIRES_ARM_NEON; |
| 3498 | GemmMicrokernelTester() |
| 3499 | .mr(6) |
| 3500 | .nr(8) |
| 3501 | .kr(1) |
| 3502 | .sr(1) |
| 3503 | .m(6) |
| 3504 | .n(8) |
| 3505 | .k(8) |
| 3506 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3507 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3508 | } |
| 3509 | |
| 3510 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, no_a_zero_point) { |
| 3511 | TEST_REQUIRES_ARM_NEON; |
| 3512 | for (size_t k = 1; k <= 40; k += 9) { |
| 3513 | GemmMicrokernelTester() |
| 3514 | .mr(6) |
| 3515 | .nr(8) |
| 3516 | .kr(1) |
| 3517 | .sr(1) |
| 3518 | .m(6) |
| 3519 | .n(8) |
| 3520 | .k(k) |
| 3521 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3522 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3523 | } |
| 3524 | } |
| 3525 | |
| 3526 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, no_b_zero_point) { |
| 3527 | TEST_REQUIRES_ARM_NEON; |
| 3528 | for (size_t k = 1; k <= 40; k += 9) { |
| 3529 | GemmMicrokernelTester() |
| 3530 | .mr(6) |
| 3531 | .nr(8) |
| 3532 | .kr(1) |
| 3533 | .sr(1) |
| 3534 | .m(6) |
| 3535 | .n(8) |
| 3536 | .k(k) |
| 3537 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3538 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3539 | } |
| 3540 | } |
| 3541 | |
| 3542 | TEST(QU8_GEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, no_zero_point) { |
| 3543 | TEST_REQUIRES_ARM_NEON; |
| 3544 | for (size_t k = 1; k <= 40; k += 9) { |
| 3545 | GemmMicrokernelTester() |
| 3546 | .mr(6) |
| 3547 | .nr(8) |
| 3548 | .kr(1) |
| 3549 | .sr(1) |
| 3550 | .m(6) |
| 3551 | .n(8) |
| 3552 | .k(k) |
| 3553 | .a_zero_point(0) |
| 3554 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3555 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3556 | } |
| 3557 | } |
| 3558 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 3559 | |
| 3560 | |
| 3561 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 3562 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8) { |
| 3563 | TEST_REQUIRES_ARM_NEON; |
| 3564 | GemmMicrokernelTester() |
| 3565 | .mr(1) |
| 3566 | .nr(16) |
| 3567 | .kr(1) |
| 3568 | .sr(1) |
| 3569 | .m(1) |
| 3570 | .n(16) |
| 3571 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3572 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3573 | } |
| 3574 | |
| 3575 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cn) { |
| 3576 | TEST_REQUIRES_ARM_NEON; |
| 3577 | GemmMicrokernelTester() |
| 3578 | .mr(1) |
| 3579 | .nr(16) |
| 3580 | .kr(1) |
| 3581 | .sr(1) |
| 3582 | .m(1) |
| 3583 | .n(16) |
| 3584 | .k(8) |
| 3585 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3586 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3587 | } |
| 3588 | |
| 3589 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_strided_a) { |
| 3590 | TEST_REQUIRES_ARM_NEON; |
| 3591 | GemmMicrokernelTester() |
| 3592 | .mr(1) |
| 3593 | .nr(16) |
| 3594 | .kr(1) |
| 3595 | .sr(1) |
| 3596 | .m(1) |
| 3597 | .n(16) |
| 3598 | .k(8) |
| 3599 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3600 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3601 | } |
| 3602 | |
| 3603 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile) { |
| 3604 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3605 | for (uint32_t n = 1; n <= 16; n++) { |
| 3606 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3607 | GemmMicrokernelTester() |
| 3608 | .mr(1) |
| 3609 | .nr(16) |
| 3610 | .kr(1) |
| 3611 | .sr(1) |
| 3612 | .m(m) |
| 3613 | .n(n) |
| 3614 | .k(8) |
| 3615 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3616 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3617 | } |
| 3618 | } |
| 3619 | } |
| 3620 | |
| 3621 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile_m) { |
| 3622 | TEST_REQUIRES_ARM_NEON; |
| 3623 | for (uint32_t m = 1; m <= 1; m++) { |
| 3624 | GemmMicrokernelTester() |
| 3625 | .mr(1) |
| 3626 | .nr(16) |
| 3627 | .kr(1) |
| 3628 | .sr(1) |
| 3629 | .m(m) |
| 3630 | .n(16) |
| 3631 | .k(8) |
| 3632 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3633 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3634 | } |
| 3635 | } |
| 3636 | |
| 3637 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile_n) { |
| 3638 | TEST_REQUIRES_ARM_NEON; |
| 3639 | for (uint32_t n = 1; n <= 16; n++) { |
| 3640 | GemmMicrokernelTester() |
| 3641 | .mr(1) |
| 3642 | .nr(16) |
| 3643 | .kr(1) |
| 3644 | .sr(1) |
| 3645 | .m(1) |
| 3646 | .n(n) |
| 3647 | .k(8) |
| 3648 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3649 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3650 | } |
| 3651 | } |
| 3652 | |
| 3653 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8) { |
| 3654 | TEST_REQUIRES_ARM_NEON; |
| 3655 | for (size_t k = 1; k < 8; k++) { |
| 3656 | GemmMicrokernelTester() |
| 3657 | .mr(1) |
| 3658 | .nr(16) |
| 3659 | .kr(1) |
| 3660 | .sr(1) |
| 3661 | .m(1) |
| 3662 | .n(16) |
| 3663 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3664 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3665 | } |
| 3666 | } |
| 3667 | |
| 3668 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8_strided_a) { |
| 3669 | TEST_REQUIRES_ARM_NEON; |
| 3670 | for (size_t k = 1; k < 8; k++) { |
| 3671 | GemmMicrokernelTester() |
| 3672 | .mr(1) |
| 3673 | .nr(16) |
| 3674 | .kr(1) |
| 3675 | .sr(1) |
| 3676 | .m(1) |
| 3677 | .n(16) |
| 3678 | .k(k) |
| 3679 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3680 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3681 | } |
| 3682 | } |
| 3683 | |
| 3684 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8_subtile) { |
| 3685 | TEST_REQUIRES_ARM_NEON; |
| 3686 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3687 | for (uint32_t n = 1; n <= 16; n++) { |
| 3688 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3689 | GemmMicrokernelTester() |
| 3690 | .mr(1) |
| 3691 | .nr(16) |
| 3692 | .kr(1) |
| 3693 | .sr(1) |
| 3694 | .m(m) |
| 3695 | .n(n) |
| 3696 | .k(k) |
| 3697 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3698 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3699 | } |
| 3700 | } |
| 3701 | } |
| 3702 | } |
| 3703 | |
| 3704 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8) { |
| 3705 | TEST_REQUIRES_ARM_NEON; |
| 3706 | for (size_t k = 9; k < 16; k++) { |
| 3707 | GemmMicrokernelTester() |
| 3708 | .mr(1) |
| 3709 | .nr(16) |
| 3710 | .kr(1) |
| 3711 | .sr(1) |
| 3712 | .m(1) |
| 3713 | .n(16) |
| 3714 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3715 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3716 | } |
| 3717 | } |
| 3718 | |
| 3719 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8_strided_a) { |
| 3720 | TEST_REQUIRES_ARM_NEON; |
| 3721 | for (size_t k = 9; k < 16; k++) { |
| 3722 | GemmMicrokernelTester() |
| 3723 | .mr(1) |
| 3724 | .nr(16) |
| 3725 | .kr(1) |
| 3726 | .sr(1) |
| 3727 | .m(1) |
| 3728 | .n(16) |
| 3729 | .k(k) |
| 3730 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3731 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3732 | } |
| 3733 | } |
| 3734 | |
| 3735 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8_subtile) { |
| 3736 | TEST_REQUIRES_ARM_NEON; |
| 3737 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3738 | for (uint32_t n = 1; n <= 16; n++) { |
| 3739 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3740 | GemmMicrokernelTester() |
| 3741 | .mr(1) |
| 3742 | .nr(16) |
| 3743 | .kr(1) |
| 3744 | .sr(1) |
| 3745 | .m(m) |
| 3746 | .n(n) |
| 3747 | .k(k) |
| 3748 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3749 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3750 | } |
| 3751 | } |
| 3752 | } |
| 3753 | } |
| 3754 | |
| 3755 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8) { |
| 3756 | TEST_REQUIRES_ARM_NEON; |
| 3757 | for (size_t k = 16; k <= 80; k += 8) { |
| 3758 | GemmMicrokernelTester() |
| 3759 | .mr(1) |
| 3760 | .nr(16) |
| 3761 | .kr(1) |
| 3762 | .sr(1) |
| 3763 | .m(1) |
| 3764 | .n(16) |
| 3765 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3766 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3767 | } |
| 3768 | } |
| 3769 | |
| 3770 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8_strided_a) { |
| 3771 | TEST_REQUIRES_ARM_NEON; |
| 3772 | for (size_t k = 16; k <= 80; k += 8) { |
| 3773 | GemmMicrokernelTester() |
| 3774 | .mr(1) |
| 3775 | .nr(16) |
| 3776 | .kr(1) |
| 3777 | .sr(1) |
| 3778 | .m(1) |
| 3779 | .n(16) |
| 3780 | .k(k) |
| 3781 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3782 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3783 | } |
| 3784 | } |
| 3785 | |
| 3786 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8_subtile) { |
| 3787 | TEST_REQUIRES_ARM_NEON; |
| 3788 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3789 | for (uint32_t n = 1; n <= 16; n++) { |
| 3790 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3791 | GemmMicrokernelTester() |
| 3792 | .mr(1) |
| 3793 | .nr(16) |
| 3794 | .kr(1) |
| 3795 | .sr(1) |
| 3796 | .m(m) |
| 3797 | .n(n) |
| 3798 | .k(k) |
| 3799 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3800 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3801 | } |
| 3802 | } |
| 3803 | } |
| 3804 | } |
| 3805 | |
| 3806 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16) { |
| 3807 | TEST_REQUIRES_ARM_NEON; |
| 3808 | for (uint32_t n = 17; n < 32; n++) { |
| 3809 | for (size_t k = 1; k <= 40; k += 9) { |
| 3810 | GemmMicrokernelTester() |
| 3811 | .mr(1) |
| 3812 | .nr(16) |
| 3813 | .kr(1) |
| 3814 | .sr(1) |
| 3815 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3816 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3817 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3818 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3819 | } |
| 3820 | } |
| 3821 | } |
| 3822 | |
| 3823 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_strided_cn) { |
| 3824 | TEST_REQUIRES_ARM_NEON; |
| 3825 | for (uint32_t n = 17; n < 32; n++) { |
| 3826 | for (size_t k = 1; k <= 40; k += 9) { |
| 3827 | GemmMicrokernelTester() |
| 3828 | .mr(1) |
| 3829 | .nr(16) |
| 3830 | .kr(1) |
| 3831 | .sr(1) |
| 3832 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3833 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3834 | .k(k) |
| 3835 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3836 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3837 | } |
| 3838 | } |
| 3839 | } |
| 3840 | |
| 3841 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_strided_a) { |
| 3842 | TEST_REQUIRES_ARM_NEON; |
| 3843 | for (uint32_t n = 17; n < 32; n++) { |
| 3844 | for (size_t k = 1; k <= 40; k += 9) { |
| 3845 | GemmMicrokernelTester() |
| 3846 | .mr(1) |
| 3847 | .nr(16) |
| 3848 | .kr(1) |
| 3849 | .sr(1) |
| 3850 | .m(1) |
| 3851 | .n(n) |
| 3852 | .k(k) |
| 3853 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3854 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3855 | } |
| 3856 | } |
| 3857 | } |
| 3858 | |
| 3859 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_subtile) { |
| 3860 | TEST_REQUIRES_ARM_NEON; |
| 3861 | for (uint32_t n = 17; n < 32; n++) { |
| 3862 | for (size_t k = 1; k <= 40; k += 9) { |
| 3863 | for (uint32_t m = 1; m <= 1; m++) { |
| 3864 | GemmMicrokernelTester() |
| 3865 | .mr(1) |
| 3866 | .nr(16) |
| 3867 | .kr(1) |
| 3868 | .sr(1) |
| 3869 | .m(m) |
| 3870 | .n(n) |
| 3871 | .k(k) |
| 3872 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3873 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3874 | } |
| 3875 | } |
| 3876 | } |
| 3877 | } |
| 3878 | |
| 3879 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16) { |
| 3880 | TEST_REQUIRES_ARM_NEON; |
| 3881 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3882 | for (size_t k = 1; k <= 40; k += 9) { |
| 3883 | GemmMicrokernelTester() |
| 3884 | .mr(1) |
| 3885 | .nr(16) |
| 3886 | .kr(1) |
| 3887 | .sr(1) |
| 3888 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3889 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3890 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3891 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3892 | } |
| 3893 | } |
| 3894 | } |
| 3895 | |
| 3896 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_strided_cn) { |
| 3897 | TEST_REQUIRES_ARM_NEON; |
| 3898 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3899 | for (size_t k = 1; k <= 40; k += 9) { |
| 3900 | GemmMicrokernelTester() |
| 3901 | .mr(1) |
| 3902 | .nr(16) |
| 3903 | .kr(1) |
| 3904 | .sr(1) |
| 3905 | .m(1) |
| 3906 | .n(n) |
| 3907 | .k(k) |
| 3908 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3909 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3910 | } |
| 3911 | } |
| 3912 | } |
| 3913 | |
| 3914 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_strided_a) { |
| 3915 | TEST_REQUIRES_ARM_NEON; |
| 3916 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3917 | for (size_t k = 1; k <= 40; k += 9) { |
| 3918 | GemmMicrokernelTester() |
| 3919 | .mr(1) |
| 3920 | .nr(16) |
| 3921 | .kr(1) |
| 3922 | .sr(1) |
| 3923 | .m(1) |
| 3924 | .n(n) |
| 3925 | .k(k) |
| 3926 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3927 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3928 | } |
| 3929 | } |
| 3930 | } |
| 3931 | |
| 3932 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_subtile) { |
| 3933 | TEST_REQUIRES_ARM_NEON; |
| 3934 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3935 | for (size_t k = 1; k <= 40; k += 9) { |
| 3936 | for (uint32_t m = 1; m <= 1; m++) { |
| 3937 | GemmMicrokernelTester() |
| 3938 | .mr(1) |
| 3939 | .nr(16) |
| 3940 | .kr(1) |
| 3941 | .sr(1) |
| 3942 | .m(m) |
| 3943 | .n(n) |
| 3944 | .k(k) |
| 3945 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3946 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3947 | } |
| 3948 | } |
| 3949 | } |
| 3950 | } |
| 3951 | |
| 3952 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cm_subtile) { |
| 3953 | TEST_REQUIRES_ARM_NEON; |
| 3954 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3955 | for (uint32_t n = 1; n <= 16; n++) { |
| 3956 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3957 | GemmMicrokernelTester() |
| 3958 | .mr(1) |
| 3959 | .nr(16) |
| 3960 | .kr(1) |
| 3961 | .sr(1) |
| 3962 | .m(m) |
| 3963 | .n(n) |
| 3964 | .k(k) |
| 3965 | .cm_stride(19) |
| 3966 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3967 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3968 | } |
| 3969 | } |
| 3970 | } |
| 3971 | } |
| 3972 | |
| 3973 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, qmin) { |
| 3974 | TEST_REQUIRES_ARM_NEON; |
| 3975 | GemmMicrokernelTester() |
| 3976 | .mr(1) |
| 3977 | .nr(16) |
| 3978 | .kr(1) |
| 3979 | .sr(1) |
| 3980 | .m(1) |
| 3981 | .n(16) |
| 3982 | .k(8) |
| 3983 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3984 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3985 | } |
| 3986 | |
| 3987 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, qmax) { |
| 3988 | TEST_REQUIRES_ARM_NEON; |
| 3989 | GemmMicrokernelTester() |
| 3990 | .mr(1) |
| 3991 | .nr(16) |
| 3992 | .kr(1) |
| 3993 | .sr(1) |
| 3994 | .m(1) |
| 3995 | .n(16) |
| 3996 | .k(8) |
| 3997 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3998 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 3999 | } |
| 4000 | |
| 4001 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cm) { |
| 4002 | TEST_REQUIRES_ARM_NEON; |
| 4003 | GemmMicrokernelTester() |
| 4004 | .mr(1) |
| 4005 | .nr(16) |
| 4006 | .kr(1) |
| 4007 | .sr(1) |
| 4008 | .m(1) |
| 4009 | .n(16) |
| 4010 | .k(8) |
| 4011 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4012 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4013 | } |
| 4014 | |
| 4015 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, no_a_zero_point) { |
| 4016 | TEST_REQUIRES_ARM_NEON; |
| 4017 | for (size_t k = 1; k <= 40; k += 9) { |
| 4018 | GemmMicrokernelTester() |
| 4019 | .mr(1) |
| 4020 | .nr(16) |
| 4021 | .kr(1) |
| 4022 | .sr(1) |
| 4023 | .m(1) |
| 4024 | .n(16) |
| 4025 | .k(k) |
| 4026 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4027 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4028 | } |
| 4029 | } |
| 4030 | |
| 4031 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, no_b_zero_point) { |
| 4032 | TEST_REQUIRES_ARM_NEON; |
| 4033 | for (size_t k = 1; k <= 40; k += 9) { |
| 4034 | GemmMicrokernelTester() |
| 4035 | .mr(1) |
| 4036 | .nr(16) |
| 4037 | .kr(1) |
| 4038 | .sr(1) |
| 4039 | .m(1) |
| 4040 | .n(16) |
| 4041 | .k(k) |
| 4042 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4043 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4044 | } |
| 4045 | } |
| 4046 | |
| 4047 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, no_zero_point) { |
| 4048 | TEST_REQUIRES_ARM_NEON; |
| 4049 | for (size_t k = 1; k <= 40; k += 9) { |
| 4050 | GemmMicrokernelTester() |
| 4051 | .mr(1) |
| 4052 | .nr(16) |
| 4053 | .kr(1) |
| 4054 | .sr(1) |
| 4055 | .m(1) |
| 4056 | .n(16) |
| 4057 | .k(k) |
| 4058 | .a_zero_point(0) |
| 4059 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4060 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4061 | } |
| 4062 | } |
| 4063 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4064 | |
| 4065 | |
| 4066 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4067 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_eq_8) { |
| 4068 | TEST_REQUIRES_ARM_NEON; |
| 4069 | GemmMicrokernelTester() |
| 4070 | .mr(2) |
| 4071 | .nr(16) |
| 4072 | .kr(1) |
| 4073 | .sr(1) |
| 4074 | .m(2) |
| 4075 | .n(16) |
| 4076 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4077 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4078 | } |
| 4079 | |
| 4080 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, strided_cn) { |
| 4081 | TEST_REQUIRES_ARM_NEON; |
| 4082 | GemmMicrokernelTester() |
| 4083 | .mr(2) |
| 4084 | .nr(16) |
| 4085 | .kr(1) |
| 4086 | .sr(1) |
| 4087 | .m(2) |
| 4088 | .n(16) |
| 4089 | .k(8) |
| 4090 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4091 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4092 | } |
| 4093 | |
| 4094 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_eq_8_strided_a) { |
| 4095 | TEST_REQUIRES_ARM_NEON; |
| 4096 | GemmMicrokernelTester() |
| 4097 | .mr(2) |
| 4098 | .nr(16) |
| 4099 | .kr(1) |
| 4100 | .sr(1) |
| 4101 | .m(2) |
| 4102 | .n(16) |
| 4103 | .k(8) |
| 4104 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4105 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4106 | } |
| 4107 | |
| 4108 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_eq_8_subtile) { |
| 4109 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4110 | for (uint32_t n = 1; n <= 16; n++) { |
| 4111 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4112 | GemmMicrokernelTester() |
| 4113 | .mr(2) |
| 4114 | .nr(16) |
| 4115 | .kr(1) |
| 4116 | .sr(1) |
| 4117 | .m(m) |
| 4118 | .n(n) |
| 4119 | .k(8) |
| 4120 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4121 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4122 | } |
| 4123 | } |
| 4124 | } |
| 4125 | |
| 4126 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_eq_8_subtile_m) { |
| 4127 | TEST_REQUIRES_ARM_NEON; |
| 4128 | for (uint32_t m = 1; m <= 2; m++) { |
| 4129 | GemmMicrokernelTester() |
| 4130 | .mr(2) |
| 4131 | .nr(16) |
| 4132 | .kr(1) |
| 4133 | .sr(1) |
| 4134 | .m(m) |
| 4135 | .n(16) |
| 4136 | .k(8) |
| 4137 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4138 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4139 | } |
| 4140 | } |
| 4141 | |
| 4142 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_eq_8_subtile_n) { |
| 4143 | TEST_REQUIRES_ARM_NEON; |
| 4144 | for (uint32_t n = 1; n <= 16; n++) { |
| 4145 | GemmMicrokernelTester() |
| 4146 | .mr(2) |
| 4147 | .nr(16) |
| 4148 | .kr(1) |
| 4149 | .sr(1) |
| 4150 | .m(2) |
| 4151 | .n(n) |
| 4152 | .k(8) |
| 4153 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4154 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4155 | } |
| 4156 | } |
| 4157 | |
| 4158 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_lt_8) { |
| 4159 | TEST_REQUIRES_ARM_NEON; |
| 4160 | for (size_t k = 1; k < 8; k++) { |
| 4161 | GemmMicrokernelTester() |
| 4162 | .mr(2) |
| 4163 | .nr(16) |
| 4164 | .kr(1) |
| 4165 | .sr(1) |
| 4166 | .m(2) |
| 4167 | .n(16) |
| 4168 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4169 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4170 | } |
| 4171 | } |
| 4172 | |
| 4173 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_lt_8_strided_a) { |
| 4174 | TEST_REQUIRES_ARM_NEON; |
| 4175 | for (size_t k = 1; k < 8; k++) { |
| 4176 | GemmMicrokernelTester() |
| 4177 | .mr(2) |
| 4178 | .nr(16) |
| 4179 | .kr(1) |
| 4180 | .sr(1) |
| 4181 | .m(2) |
| 4182 | .n(16) |
| 4183 | .k(k) |
| 4184 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4185 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4186 | } |
| 4187 | } |
| 4188 | |
| 4189 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_lt_8_subtile) { |
| 4190 | TEST_REQUIRES_ARM_NEON; |
| 4191 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4192 | for (uint32_t n = 1; n <= 16; n++) { |
| 4193 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4194 | GemmMicrokernelTester() |
| 4195 | .mr(2) |
| 4196 | .nr(16) |
| 4197 | .kr(1) |
| 4198 | .sr(1) |
| 4199 | .m(m) |
| 4200 | .n(n) |
| 4201 | .k(k) |
| 4202 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4203 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4204 | } |
| 4205 | } |
| 4206 | } |
| 4207 | } |
| 4208 | |
| 4209 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_gt_8) { |
| 4210 | TEST_REQUIRES_ARM_NEON; |
| 4211 | for (size_t k = 9; k < 16; k++) { |
| 4212 | GemmMicrokernelTester() |
| 4213 | .mr(2) |
| 4214 | .nr(16) |
| 4215 | .kr(1) |
| 4216 | .sr(1) |
| 4217 | .m(2) |
| 4218 | .n(16) |
| 4219 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4220 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4221 | } |
| 4222 | } |
| 4223 | |
| 4224 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_gt_8_strided_a) { |
| 4225 | TEST_REQUIRES_ARM_NEON; |
| 4226 | for (size_t k = 9; k < 16; k++) { |
| 4227 | GemmMicrokernelTester() |
| 4228 | .mr(2) |
| 4229 | .nr(16) |
| 4230 | .kr(1) |
| 4231 | .sr(1) |
| 4232 | .m(2) |
| 4233 | .n(16) |
| 4234 | .k(k) |
| 4235 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4236 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4237 | } |
| 4238 | } |
| 4239 | |
| 4240 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_gt_8_subtile) { |
| 4241 | TEST_REQUIRES_ARM_NEON; |
| 4242 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4243 | for (uint32_t n = 1; n <= 16; n++) { |
| 4244 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4245 | GemmMicrokernelTester() |
| 4246 | .mr(2) |
| 4247 | .nr(16) |
| 4248 | .kr(1) |
| 4249 | .sr(1) |
| 4250 | .m(m) |
| 4251 | .n(n) |
| 4252 | .k(k) |
| 4253 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4254 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4255 | } |
| 4256 | } |
| 4257 | } |
| 4258 | } |
| 4259 | |
| 4260 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_div_8) { |
| 4261 | TEST_REQUIRES_ARM_NEON; |
| 4262 | for (size_t k = 16; k <= 80; k += 8) { |
| 4263 | GemmMicrokernelTester() |
| 4264 | .mr(2) |
| 4265 | .nr(16) |
| 4266 | .kr(1) |
| 4267 | .sr(1) |
| 4268 | .m(2) |
| 4269 | .n(16) |
| 4270 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4271 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4272 | } |
| 4273 | } |
| 4274 | |
| 4275 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_div_8_strided_a) { |
| 4276 | TEST_REQUIRES_ARM_NEON; |
| 4277 | for (size_t k = 16; k <= 80; k += 8) { |
| 4278 | GemmMicrokernelTester() |
| 4279 | .mr(2) |
| 4280 | .nr(16) |
| 4281 | .kr(1) |
| 4282 | .sr(1) |
| 4283 | .m(2) |
| 4284 | .n(16) |
| 4285 | .k(k) |
| 4286 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4287 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4288 | } |
| 4289 | } |
| 4290 | |
| 4291 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, k_div_8_subtile) { |
| 4292 | TEST_REQUIRES_ARM_NEON; |
| 4293 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4294 | for (uint32_t n = 1; n <= 16; n++) { |
| 4295 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4296 | GemmMicrokernelTester() |
| 4297 | .mr(2) |
| 4298 | .nr(16) |
| 4299 | .kr(1) |
| 4300 | .sr(1) |
| 4301 | .m(m) |
| 4302 | .n(n) |
| 4303 | .k(k) |
| 4304 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4305 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4306 | } |
| 4307 | } |
| 4308 | } |
| 4309 | } |
| 4310 | |
| 4311 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_gt_16) { |
| 4312 | TEST_REQUIRES_ARM_NEON; |
| 4313 | for (uint32_t n = 17; n < 32; n++) { |
| 4314 | for (size_t k = 1; k <= 40; k += 9) { |
| 4315 | GemmMicrokernelTester() |
| 4316 | .mr(2) |
| 4317 | .nr(16) |
| 4318 | .kr(1) |
| 4319 | .sr(1) |
| 4320 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4321 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4322 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4323 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4324 | } |
| 4325 | } |
| 4326 | } |
| 4327 | |
| 4328 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_gt_16_strided_cn) { |
| 4329 | TEST_REQUIRES_ARM_NEON; |
| 4330 | for (uint32_t n = 17; n < 32; n++) { |
| 4331 | for (size_t k = 1; k <= 40; k += 9) { |
| 4332 | GemmMicrokernelTester() |
| 4333 | .mr(2) |
| 4334 | .nr(16) |
| 4335 | .kr(1) |
| 4336 | .sr(1) |
| 4337 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4338 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4339 | .k(k) |
| 4340 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4341 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4342 | } |
| 4343 | } |
| 4344 | } |
| 4345 | |
| 4346 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_gt_16_strided_a) { |
| 4347 | TEST_REQUIRES_ARM_NEON; |
| 4348 | for (uint32_t n = 17; n < 32; n++) { |
| 4349 | for (size_t k = 1; k <= 40; k += 9) { |
| 4350 | GemmMicrokernelTester() |
| 4351 | .mr(2) |
| 4352 | .nr(16) |
| 4353 | .kr(1) |
| 4354 | .sr(1) |
| 4355 | .m(2) |
| 4356 | .n(n) |
| 4357 | .k(k) |
| 4358 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4359 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4360 | } |
| 4361 | } |
| 4362 | } |
| 4363 | |
| 4364 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_gt_16_subtile) { |
| 4365 | TEST_REQUIRES_ARM_NEON; |
| 4366 | for (uint32_t n = 17; n < 32; n++) { |
| 4367 | for (size_t k = 1; k <= 40; k += 9) { |
| 4368 | for (uint32_t m = 1; m <= 2; m++) { |
| 4369 | GemmMicrokernelTester() |
| 4370 | .mr(2) |
| 4371 | .nr(16) |
| 4372 | .kr(1) |
| 4373 | .sr(1) |
| 4374 | .m(m) |
| 4375 | .n(n) |
| 4376 | .k(k) |
| 4377 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4378 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4379 | } |
| 4380 | } |
| 4381 | } |
| 4382 | } |
| 4383 | |
| 4384 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_div_16) { |
| 4385 | TEST_REQUIRES_ARM_NEON; |
| 4386 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4387 | for (size_t k = 1; k <= 40; k += 9) { |
| 4388 | GemmMicrokernelTester() |
| 4389 | .mr(2) |
| 4390 | .nr(16) |
| 4391 | .kr(1) |
| 4392 | .sr(1) |
| 4393 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4394 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4395 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4396 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4397 | } |
| 4398 | } |
| 4399 | } |
| 4400 | |
| 4401 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_div_16_strided_cn) { |
| 4402 | TEST_REQUIRES_ARM_NEON; |
| 4403 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4404 | for (size_t k = 1; k <= 40; k += 9) { |
| 4405 | GemmMicrokernelTester() |
| 4406 | .mr(2) |
| 4407 | .nr(16) |
| 4408 | .kr(1) |
| 4409 | .sr(1) |
| 4410 | .m(2) |
| 4411 | .n(n) |
| 4412 | .k(k) |
| 4413 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4414 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4415 | } |
| 4416 | } |
| 4417 | } |
| 4418 | |
| 4419 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_div_16_strided_a) { |
| 4420 | TEST_REQUIRES_ARM_NEON; |
| 4421 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4422 | for (size_t k = 1; k <= 40; k += 9) { |
| 4423 | GemmMicrokernelTester() |
| 4424 | .mr(2) |
| 4425 | .nr(16) |
| 4426 | .kr(1) |
| 4427 | .sr(1) |
| 4428 | .m(2) |
| 4429 | .n(n) |
| 4430 | .k(k) |
| 4431 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4432 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4433 | } |
| 4434 | } |
| 4435 | } |
| 4436 | |
| 4437 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, n_div_16_subtile) { |
| 4438 | TEST_REQUIRES_ARM_NEON; |
| 4439 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4440 | for (size_t k = 1; k <= 40; k += 9) { |
| 4441 | for (uint32_t m = 1; m <= 2; m++) { |
| 4442 | GemmMicrokernelTester() |
| 4443 | .mr(2) |
| 4444 | .nr(16) |
| 4445 | .kr(1) |
| 4446 | .sr(1) |
| 4447 | .m(m) |
| 4448 | .n(n) |
| 4449 | .k(k) |
| 4450 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4451 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4452 | } |
| 4453 | } |
| 4454 | } |
| 4455 | } |
| 4456 | |
| 4457 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, strided_cm_subtile) { |
| 4458 | TEST_REQUIRES_ARM_NEON; |
| 4459 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4460 | for (uint32_t n = 1; n <= 16; n++) { |
| 4461 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4462 | GemmMicrokernelTester() |
| 4463 | .mr(2) |
| 4464 | .nr(16) |
| 4465 | .kr(1) |
| 4466 | .sr(1) |
| 4467 | .m(m) |
| 4468 | .n(n) |
| 4469 | .k(k) |
| 4470 | .cm_stride(19) |
| 4471 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4472 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4473 | } |
| 4474 | } |
| 4475 | } |
| 4476 | } |
| 4477 | |
| 4478 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, qmin) { |
| 4479 | TEST_REQUIRES_ARM_NEON; |
| 4480 | GemmMicrokernelTester() |
| 4481 | .mr(2) |
| 4482 | .nr(16) |
| 4483 | .kr(1) |
| 4484 | .sr(1) |
| 4485 | .m(2) |
| 4486 | .n(16) |
| 4487 | .k(8) |
| 4488 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4489 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4490 | } |
| 4491 | |
| 4492 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, qmax) { |
| 4493 | TEST_REQUIRES_ARM_NEON; |
| 4494 | GemmMicrokernelTester() |
| 4495 | .mr(2) |
| 4496 | .nr(16) |
| 4497 | .kr(1) |
| 4498 | .sr(1) |
| 4499 | .m(2) |
| 4500 | .n(16) |
| 4501 | .k(8) |
| 4502 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4503 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4504 | } |
| 4505 | |
| 4506 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, strided_cm) { |
| 4507 | TEST_REQUIRES_ARM_NEON; |
| 4508 | GemmMicrokernelTester() |
| 4509 | .mr(2) |
| 4510 | .nr(16) |
| 4511 | .kr(1) |
| 4512 | .sr(1) |
| 4513 | .m(2) |
| 4514 | .n(16) |
| 4515 | .k(8) |
| 4516 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4517 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4518 | } |
| 4519 | |
| 4520 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, no_a_zero_point) { |
| 4521 | TEST_REQUIRES_ARM_NEON; |
| 4522 | for (size_t k = 1; k <= 40; k += 9) { |
| 4523 | GemmMicrokernelTester() |
| 4524 | .mr(2) |
| 4525 | .nr(16) |
| 4526 | .kr(1) |
| 4527 | .sr(1) |
| 4528 | .m(2) |
| 4529 | .n(16) |
| 4530 | .k(k) |
| 4531 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4532 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4533 | } |
| 4534 | } |
| 4535 | |
| 4536 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, no_b_zero_point) { |
| 4537 | TEST_REQUIRES_ARM_NEON; |
| 4538 | for (size_t k = 1; k <= 40; k += 9) { |
| 4539 | GemmMicrokernelTester() |
| 4540 | .mr(2) |
| 4541 | .nr(16) |
| 4542 | .kr(1) |
| 4543 | .sr(1) |
| 4544 | .m(2) |
| 4545 | .n(16) |
| 4546 | .k(k) |
| 4547 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4548 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4549 | } |
| 4550 | } |
| 4551 | |
| 4552 | TEST(QU8_GEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE, no_zero_point) { |
| 4553 | TEST_REQUIRES_ARM_NEON; |
| 4554 | for (size_t k = 1; k <= 40; k += 9) { |
| 4555 | GemmMicrokernelTester() |
| 4556 | .mr(2) |
| 4557 | .nr(16) |
| 4558 | .kr(1) |
| 4559 | .sr(1) |
| 4560 | .m(2) |
| 4561 | .n(16) |
| 4562 | .k(k) |
| 4563 | .a_zero_point(0) |
| 4564 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4565 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4566 | } |
| 4567 | } |
| 4568 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4569 | |
| 4570 | |
| 4571 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4572 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_eq_8) { |
| 4573 | TEST_REQUIRES_ARM_NEON; |
| 4574 | GemmMicrokernelTester() |
| 4575 | .mr(3) |
| 4576 | .nr(16) |
| 4577 | .kr(1) |
| 4578 | .sr(1) |
| 4579 | .m(3) |
| 4580 | .n(16) |
| 4581 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4582 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4583 | } |
| 4584 | |
| 4585 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, strided_cn) { |
| 4586 | TEST_REQUIRES_ARM_NEON; |
| 4587 | GemmMicrokernelTester() |
| 4588 | .mr(3) |
| 4589 | .nr(16) |
| 4590 | .kr(1) |
| 4591 | .sr(1) |
| 4592 | .m(3) |
| 4593 | .n(16) |
| 4594 | .k(8) |
| 4595 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4596 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4597 | } |
| 4598 | |
| 4599 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_eq_8_strided_a) { |
| 4600 | TEST_REQUIRES_ARM_NEON; |
| 4601 | GemmMicrokernelTester() |
| 4602 | .mr(3) |
| 4603 | .nr(16) |
| 4604 | .kr(1) |
| 4605 | .sr(1) |
| 4606 | .m(3) |
| 4607 | .n(16) |
| 4608 | .k(8) |
| 4609 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4610 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4611 | } |
| 4612 | |
| 4613 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_eq_8_subtile) { |
| 4614 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4615 | for (uint32_t n = 1; n <= 16; n++) { |
| 4616 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4617 | GemmMicrokernelTester() |
| 4618 | .mr(3) |
| 4619 | .nr(16) |
| 4620 | .kr(1) |
| 4621 | .sr(1) |
| 4622 | .m(m) |
| 4623 | .n(n) |
| 4624 | .k(8) |
| 4625 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4626 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4627 | } |
| 4628 | } |
| 4629 | } |
| 4630 | |
| 4631 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_eq_8_subtile_m) { |
| 4632 | TEST_REQUIRES_ARM_NEON; |
| 4633 | for (uint32_t m = 1; m <= 3; m++) { |
| 4634 | GemmMicrokernelTester() |
| 4635 | .mr(3) |
| 4636 | .nr(16) |
| 4637 | .kr(1) |
| 4638 | .sr(1) |
| 4639 | .m(m) |
| 4640 | .n(16) |
| 4641 | .k(8) |
| 4642 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4643 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4644 | } |
| 4645 | } |
| 4646 | |
| 4647 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_eq_8_subtile_n) { |
| 4648 | TEST_REQUIRES_ARM_NEON; |
| 4649 | for (uint32_t n = 1; n <= 16; n++) { |
| 4650 | GemmMicrokernelTester() |
| 4651 | .mr(3) |
| 4652 | .nr(16) |
| 4653 | .kr(1) |
| 4654 | .sr(1) |
| 4655 | .m(3) |
| 4656 | .n(n) |
| 4657 | .k(8) |
| 4658 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4659 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4660 | } |
| 4661 | } |
| 4662 | |
| 4663 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_lt_8) { |
| 4664 | TEST_REQUIRES_ARM_NEON; |
| 4665 | for (size_t k = 1; k < 8; k++) { |
| 4666 | GemmMicrokernelTester() |
| 4667 | .mr(3) |
| 4668 | .nr(16) |
| 4669 | .kr(1) |
| 4670 | .sr(1) |
| 4671 | .m(3) |
| 4672 | .n(16) |
| 4673 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4674 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4675 | } |
| 4676 | } |
| 4677 | |
| 4678 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_lt_8_strided_a) { |
| 4679 | TEST_REQUIRES_ARM_NEON; |
| 4680 | for (size_t k = 1; k < 8; k++) { |
| 4681 | GemmMicrokernelTester() |
| 4682 | .mr(3) |
| 4683 | .nr(16) |
| 4684 | .kr(1) |
| 4685 | .sr(1) |
| 4686 | .m(3) |
| 4687 | .n(16) |
| 4688 | .k(k) |
| 4689 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4690 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4691 | } |
| 4692 | } |
| 4693 | |
| 4694 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_lt_8_subtile) { |
| 4695 | TEST_REQUIRES_ARM_NEON; |
| 4696 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4697 | for (uint32_t n = 1; n <= 16; n++) { |
| 4698 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4699 | GemmMicrokernelTester() |
| 4700 | .mr(3) |
| 4701 | .nr(16) |
| 4702 | .kr(1) |
| 4703 | .sr(1) |
| 4704 | .m(m) |
| 4705 | .n(n) |
| 4706 | .k(k) |
| 4707 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4708 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4709 | } |
| 4710 | } |
| 4711 | } |
| 4712 | } |
| 4713 | |
| 4714 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_gt_8) { |
| 4715 | TEST_REQUIRES_ARM_NEON; |
| 4716 | for (size_t k = 9; k < 16; k++) { |
| 4717 | GemmMicrokernelTester() |
| 4718 | .mr(3) |
| 4719 | .nr(16) |
| 4720 | .kr(1) |
| 4721 | .sr(1) |
| 4722 | .m(3) |
| 4723 | .n(16) |
| 4724 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4725 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4726 | } |
| 4727 | } |
| 4728 | |
| 4729 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_gt_8_strided_a) { |
| 4730 | TEST_REQUIRES_ARM_NEON; |
| 4731 | for (size_t k = 9; k < 16; k++) { |
| 4732 | GemmMicrokernelTester() |
| 4733 | .mr(3) |
| 4734 | .nr(16) |
| 4735 | .kr(1) |
| 4736 | .sr(1) |
| 4737 | .m(3) |
| 4738 | .n(16) |
| 4739 | .k(k) |
| 4740 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4741 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4742 | } |
| 4743 | } |
| 4744 | |
| 4745 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_gt_8_subtile) { |
| 4746 | TEST_REQUIRES_ARM_NEON; |
| 4747 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4748 | for (uint32_t n = 1; n <= 16; n++) { |
| 4749 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4750 | GemmMicrokernelTester() |
| 4751 | .mr(3) |
| 4752 | .nr(16) |
| 4753 | .kr(1) |
| 4754 | .sr(1) |
| 4755 | .m(m) |
| 4756 | .n(n) |
| 4757 | .k(k) |
| 4758 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4759 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4760 | } |
| 4761 | } |
| 4762 | } |
| 4763 | } |
| 4764 | |
| 4765 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_div_8) { |
| 4766 | TEST_REQUIRES_ARM_NEON; |
| 4767 | for (size_t k = 16; k <= 80; k += 8) { |
| 4768 | GemmMicrokernelTester() |
| 4769 | .mr(3) |
| 4770 | .nr(16) |
| 4771 | .kr(1) |
| 4772 | .sr(1) |
| 4773 | .m(3) |
| 4774 | .n(16) |
| 4775 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4776 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4777 | } |
| 4778 | } |
| 4779 | |
| 4780 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_div_8_strided_a) { |
| 4781 | TEST_REQUIRES_ARM_NEON; |
| 4782 | for (size_t k = 16; k <= 80; k += 8) { |
| 4783 | GemmMicrokernelTester() |
| 4784 | .mr(3) |
| 4785 | .nr(16) |
| 4786 | .kr(1) |
| 4787 | .sr(1) |
| 4788 | .m(3) |
| 4789 | .n(16) |
| 4790 | .k(k) |
| 4791 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4792 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4793 | } |
| 4794 | } |
| 4795 | |
| 4796 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, k_div_8_subtile) { |
| 4797 | TEST_REQUIRES_ARM_NEON; |
| 4798 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4799 | for (uint32_t n = 1; n <= 16; n++) { |
| 4800 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4801 | GemmMicrokernelTester() |
| 4802 | .mr(3) |
| 4803 | .nr(16) |
| 4804 | .kr(1) |
| 4805 | .sr(1) |
| 4806 | .m(m) |
| 4807 | .n(n) |
| 4808 | .k(k) |
| 4809 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4810 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4811 | } |
| 4812 | } |
| 4813 | } |
| 4814 | } |
| 4815 | |
| 4816 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_gt_16) { |
| 4817 | TEST_REQUIRES_ARM_NEON; |
| 4818 | for (uint32_t n = 17; n < 32; n++) { |
| 4819 | for (size_t k = 1; k <= 40; k += 9) { |
| 4820 | GemmMicrokernelTester() |
| 4821 | .mr(3) |
| 4822 | .nr(16) |
| 4823 | .kr(1) |
| 4824 | .sr(1) |
| 4825 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4826 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4827 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4828 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4829 | } |
| 4830 | } |
| 4831 | } |
| 4832 | |
| 4833 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_gt_16_strided_cn) { |
| 4834 | TEST_REQUIRES_ARM_NEON; |
| 4835 | for (uint32_t n = 17; n < 32; n++) { |
| 4836 | for (size_t k = 1; k <= 40; k += 9) { |
| 4837 | GemmMicrokernelTester() |
| 4838 | .mr(3) |
| 4839 | .nr(16) |
| 4840 | .kr(1) |
| 4841 | .sr(1) |
| 4842 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4843 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4844 | .k(k) |
| 4845 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4846 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4847 | } |
| 4848 | } |
| 4849 | } |
| 4850 | |
| 4851 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_gt_16_strided_a) { |
| 4852 | TEST_REQUIRES_ARM_NEON; |
| 4853 | for (uint32_t n = 17; n < 32; n++) { |
| 4854 | for (size_t k = 1; k <= 40; k += 9) { |
| 4855 | GemmMicrokernelTester() |
| 4856 | .mr(3) |
| 4857 | .nr(16) |
| 4858 | .kr(1) |
| 4859 | .sr(1) |
| 4860 | .m(3) |
| 4861 | .n(n) |
| 4862 | .k(k) |
| 4863 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4864 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4865 | } |
| 4866 | } |
| 4867 | } |
| 4868 | |
| 4869 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_gt_16_subtile) { |
| 4870 | TEST_REQUIRES_ARM_NEON; |
| 4871 | for (uint32_t n = 17; n < 32; n++) { |
| 4872 | for (size_t k = 1; k <= 40; k += 9) { |
| 4873 | for (uint32_t m = 1; m <= 3; m++) { |
| 4874 | GemmMicrokernelTester() |
| 4875 | .mr(3) |
| 4876 | .nr(16) |
| 4877 | .kr(1) |
| 4878 | .sr(1) |
| 4879 | .m(m) |
| 4880 | .n(n) |
| 4881 | .k(k) |
| 4882 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4883 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4884 | } |
| 4885 | } |
| 4886 | } |
| 4887 | } |
| 4888 | |
| 4889 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_div_16) { |
| 4890 | TEST_REQUIRES_ARM_NEON; |
| 4891 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4892 | for (size_t k = 1; k <= 40; k += 9) { |
| 4893 | GemmMicrokernelTester() |
| 4894 | .mr(3) |
| 4895 | .nr(16) |
| 4896 | .kr(1) |
| 4897 | .sr(1) |
| 4898 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4899 | .n(n) |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4900 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4901 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4902 | } |
| 4903 | } |
| 4904 | } |
| 4905 | |
| 4906 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_div_16_strided_cn) { |
| 4907 | TEST_REQUIRES_ARM_NEON; |
| 4908 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4909 | for (size_t k = 1; k <= 40; k += 9) { |
| 4910 | GemmMicrokernelTester() |
| 4911 | .mr(3) |
| 4912 | .nr(16) |
| 4913 | .kr(1) |
| 4914 | .sr(1) |
| 4915 | .m(3) |
| 4916 | .n(n) |
| 4917 | .k(k) |
| 4918 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4919 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4920 | } |
| 4921 | } |
| 4922 | } |
| 4923 | |
| 4924 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_div_16_strided_a) { |
| 4925 | TEST_REQUIRES_ARM_NEON; |
| 4926 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4927 | for (size_t k = 1; k <= 40; k += 9) { |
| 4928 | GemmMicrokernelTester() |
| 4929 | .mr(3) |
| 4930 | .nr(16) |
| 4931 | .kr(1) |
| 4932 | .sr(1) |
| 4933 | .m(3) |
| 4934 | .n(n) |
| 4935 | .k(k) |
| 4936 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4937 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4938 | } |
| 4939 | } |
| 4940 | } |
| 4941 | |
| 4942 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, n_div_16_subtile) { |
| 4943 | TEST_REQUIRES_ARM_NEON; |
| 4944 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4945 | for (size_t k = 1; k <= 40; k += 9) { |
| 4946 | for (uint32_t m = 1; m <= 3; m++) { |
| 4947 | GemmMicrokernelTester() |
| 4948 | .mr(3) |
| 4949 | .nr(16) |
| 4950 | .kr(1) |
| 4951 | .sr(1) |
| 4952 | .m(m) |
| 4953 | .n(n) |
| 4954 | .k(k) |
| 4955 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4956 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4957 | } |
| 4958 | } |
| 4959 | } |
| 4960 | } |
| 4961 | |
| 4962 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, strided_cm_subtile) { |
| 4963 | TEST_REQUIRES_ARM_NEON; |
| 4964 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4965 | for (uint32_t n = 1; n <= 16; n++) { |
| 4966 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4967 | GemmMicrokernelTester() |
| 4968 | .mr(3) |
| 4969 | .nr(16) |
| 4970 | .kr(1) |
| 4971 | .sr(1) |
| 4972 | .m(m) |
| 4973 | .n(n) |
| 4974 | .k(k) |
| 4975 | .cm_stride(19) |
| 4976 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4977 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4978 | } |
| 4979 | } |
| 4980 | } |
| 4981 | } |
| 4982 | |
| 4983 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, qmin) { |
| 4984 | TEST_REQUIRES_ARM_NEON; |
| 4985 | GemmMicrokernelTester() |
| 4986 | .mr(3) |
| 4987 | .nr(16) |
| 4988 | .kr(1) |
| 4989 | .sr(1) |
| 4990 | .m(3) |
| 4991 | .n(16) |
| 4992 | .k(8) |
| 4993 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4994 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 4995 | } |
| 4996 | |
| 4997 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, qmax) { |
| 4998 | TEST_REQUIRES_ARM_NEON; |
| 4999 | GemmMicrokernelTester() |
| 5000 | .mr(3) |
| 5001 | .nr(16) |
| 5002 | .kr(1) |
| 5003 | .sr(1) |
| 5004 | .m(3) |
| 5005 | .n(16) |
| 5006 | .k(8) |
| 5007 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5008 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 5009 | } |
| 5010 | |
| 5011 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, strided_cm) { |
| 5012 | TEST_REQUIRES_ARM_NEON; |
| 5013 | GemmMicrokernelTester() |
| 5014 | .mr(3) |
| 5015 | .nr(16) |
| 5016 | .kr(1) |
| 5017 | .sr(1) |
| 5018 | .m(3) |
| 5019 | .n(16) |
| 5020 | .k(8) |
| 5021 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5022 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 5023 | } |
| 5024 | |
| 5025 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, no_a_zero_point) { |
| 5026 | TEST_REQUIRES_ARM_NEON; |
| 5027 | for (size_t k = 1; k <= 40; k += 9) { |
| 5028 | GemmMicrokernelTester() |
| 5029 | .mr(3) |
| 5030 | .nr(16) |
| 5031 | .kr(1) |
| 5032 | .sr(1) |
| 5033 | .m(3) |
| 5034 | .n(16) |
| 5035 | .k(k) |
| 5036 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5037 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 5038 | } |
| 5039 | } |
| 5040 | |
| 5041 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, no_b_zero_point) { |
| 5042 | TEST_REQUIRES_ARM_NEON; |
| 5043 | for (size_t k = 1; k <= 40; k += 9) { |
| 5044 | GemmMicrokernelTester() |
| 5045 | .mr(3) |
| 5046 | .nr(16) |
| 5047 | .kr(1) |
| 5048 | .sr(1) |
| 5049 | .m(3) |
| 5050 | .n(16) |
| 5051 | .k(k) |
| 5052 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5053 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 5054 | } |
| 5055 | } |
| 5056 | |
| 5057 | TEST(QU8_GEMM_MINMAX_RNDNU_3X16__NEON_MLAL_LANE, no_zero_point) { |
| 5058 | TEST_REQUIRES_ARM_NEON; |
| 5059 | for (size_t k = 1; k <= 40; k += 9) { |
| 5060 | GemmMicrokernelTester() |
| 5061 | .mr(3) |
| 5062 | .nr(16) |
| 5063 | .kr(1) |
| 5064 | .sr(1) |
| 5065 | .m(3) |
| 5066 | .n(16) |
| 5067 | .k(k) |
| 5068 | .a_zero_point(0) |
| 5069 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5070 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x16__neon_mlal_lane, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | d5a5333 | 2022-01-10 03:44:40 -0800 | [diff] [blame] | 5071 | } |
| 5072 | } |
| 5073 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 5074 | |
| 5075 | |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 5076 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5077 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16) { |
| 5078 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5079 | GemmMicrokernelTester() |
| 5080 | .mr(4) |
| 5081 | .nr(16) |
| 5082 | .kr(4) |
| 5083 | .sr(1) |
| 5084 | .m(4) |
| 5085 | .n(16) |
| 5086 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5087 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5088 | } |
| 5089 | |
| 5090 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cn) { |
| 5091 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5092 | GemmMicrokernelTester() |
| 5093 | .mr(4) |
| 5094 | .nr(16) |
| 5095 | .kr(4) |
| 5096 | .sr(1) |
| 5097 | .m(4) |
| 5098 | .n(16) |
| 5099 | .k(16) |
| 5100 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5101 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5102 | } |
| 5103 | |
| 5104 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_strided_a) { |
| 5105 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5106 | GemmMicrokernelTester() |
| 5107 | .mr(4) |
| 5108 | .nr(16) |
| 5109 | .kr(4) |
| 5110 | .sr(1) |
| 5111 | .m(4) |
| 5112 | .n(16) |
| 5113 | .k(16) |
| 5114 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5115 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5116 | } |
| 5117 | |
| 5118 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile) { |
| 5119 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5120 | for (uint32_t n = 1; n <= 16; n++) { |
| 5121 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5122 | GemmMicrokernelTester() |
| 5123 | .mr(4) |
| 5124 | .nr(16) |
| 5125 | .kr(4) |
| 5126 | .sr(1) |
| 5127 | .m(m) |
| 5128 | .n(n) |
| 5129 | .k(16) |
| 5130 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5131 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5132 | } |
| 5133 | } |
| 5134 | } |
| 5135 | |
| 5136 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile_m) { |
| 5137 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5138 | for (uint32_t m = 1; m <= 4; m++) { |
| 5139 | GemmMicrokernelTester() |
| 5140 | .mr(4) |
| 5141 | .nr(16) |
| 5142 | .kr(4) |
| 5143 | .sr(1) |
| 5144 | .m(m) |
| 5145 | .n(16) |
| 5146 | .k(16) |
| 5147 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5148 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5149 | } |
| 5150 | } |
| 5151 | |
| 5152 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile_n) { |
| 5153 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5154 | for (uint32_t n = 1; n <= 16; n++) { |
| 5155 | GemmMicrokernelTester() |
| 5156 | .mr(4) |
| 5157 | .nr(16) |
| 5158 | .kr(4) |
| 5159 | .sr(1) |
| 5160 | .m(4) |
| 5161 | .n(n) |
| 5162 | .k(16) |
| 5163 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5164 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5165 | } |
| 5166 | } |
| 5167 | |
| 5168 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16) { |
| 5169 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5170 | for (size_t k = 1; k < 16; k++) { |
| 5171 | GemmMicrokernelTester() |
| 5172 | .mr(4) |
| 5173 | .nr(16) |
| 5174 | .kr(4) |
| 5175 | .sr(1) |
| 5176 | .m(4) |
| 5177 | .n(16) |
| 5178 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5179 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5180 | } |
| 5181 | } |
| 5182 | |
| 5183 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16_strided_a) { |
| 5184 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5185 | for (size_t k = 1; k < 16; k++) { |
| 5186 | GemmMicrokernelTester() |
| 5187 | .mr(4) |
| 5188 | .nr(16) |
| 5189 | .kr(4) |
| 5190 | .sr(1) |
| 5191 | .m(4) |
| 5192 | .n(16) |
| 5193 | .k(k) |
| 5194 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5195 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5196 | } |
| 5197 | } |
| 5198 | |
| 5199 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16_subtile) { |
| 5200 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5201 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5202 | for (uint32_t n = 1; n <= 16; n++) { |
| 5203 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5204 | GemmMicrokernelTester() |
| 5205 | .mr(4) |
| 5206 | .nr(16) |
| 5207 | .kr(4) |
| 5208 | .sr(1) |
| 5209 | .m(m) |
| 5210 | .n(n) |
| 5211 | .k(k) |
| 5212 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5213 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5214 | } |
| 5215 | } |
| 5216 | } |
| 5217 | } |
| 5218 | |
| 5219 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16) { |
| 5220 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5221 | for (size_t k = 17; k < 32; k++) { |
| 5222 | GemmMicrokernelTester() |
| 5223 | .mr(4) |
| 5224 | .nr(16) |
| 5225 | .kr(4) |
| 5226 | .sr(1) |
| 5227 | .m(4) |
| 5228 | .n(16) |
| 5229 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5230 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5231 | } |
| 5232 | } |
| 5233 | |
| 5234 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16_strided_a) { |
| 5235 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5236 | for (size_t k = 17; k < 32; k++) { |
| 5237 | GemmMicrokernelTester() |
| 5238 | .mr(4) |
| 5239 | .nr(16) |
| 5240 | .kr(4) |
| 5241 | .sr(1) |
| 5242 | .m(4) |
| 5243 | .n(16) |
| 5244 | .k(k) |
| 5245 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5246 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5247 | } |
| 5248 | } |
| 5249 | |
| 5250 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16_subtile) { |
| 5251 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5252 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5253 | for (uint32_t n = 1; n <= 16; n++) { |
| 5254 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5255 | GemmMicrokernelTester() |
| 5256 | .mr(4) |
| 5257 | .nr(16) |
| 5258 | .kr(4) |
| 5259 | .sr(1) |
| 5260 | .m(m) |
| 5261 | .n(n) |
| 5262 | .k(k) |
| 5263 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5264 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5265 | } |
| 5266 | } |
| 5267 | } |
| 5268 | } |
| 5269 | |
| 5270 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16) { |
| 5271 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5272 | for (size_t k = 32; k <= 160; k += 16) { |
| 5273 | GemmMicrokernelTester() |
| 5274 | .mr(4) |
| 5275 | .nr(16) |
| 5276 | .kr(4) |
| 5277 | .sr(1) |
| 5278 | .m(4) |
| 5279 | .n(16) |
| 5280 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5281 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5282 | } |
| 5283 | } |
| 5284 | |
| 5285 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16_strided_a) { |
| 5286 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5287 | for (size_t k = 32; k <= 160; k += 16) { |
| 5288 | GemmMicrokernelTester() |
| 5289 | .mr(4) |
| 5290 | .nr(16) |
| 5291 | .kr(4) |
| 5292 | .sr(1) |
| 5293 | .m(4) |
| 5294 | .n(16) |
| 5295 | .k(k) |
| 5296 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5297 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5298 | } |
| 5299 | } |
| 5300 | |
| 5301 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16_subtile) { |
| 5302 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5303 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5304 | for (uint32_t n = 1; n <= 16; n++) { |
| 5305 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5306 | GemmMicrokernelTester() |
| 5307 | .mr(4) |
| 5308 | .nr(16) |
| 5309 | .kr(4) |
| 5310 | .sr(1) |
| 5311 | .m(m) |
| 5312 | .n(n) |
| 5313 | .k(k) |
| 5314 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5315 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5316 | } |
| 5317 | } |
| 5318 | } |
| 5319 | } |
| 5320 | |
| 5321 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16) { |
| 5322 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5323 | for (uint32_t n = 17; n < 32; n++) { |
| 5324 | for (size_t k = 1; k <= 80; k += 17) { |
| 5325 | GemmMicrokernelTester() |
| 5326 | .mr(4) |
| 5327 | .nr(16) |
| 5328 | .kr(4) |
| 5329 | .sr(1) |
| 5330 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5331 | .n(n) |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5332 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5333 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5334 | } |
| 5335 | } |
| 5336 | } |
| 5337 | |
| 5338 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_strided_cn) { |
| 5339 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5340 | for (uint32_t n = 17; n < 32; n++) { |
| 5341 | for (size_t k = 1; k <= 80; k += 17) { |
| 5342 | GemmMicrokernelTester() |
| 5343 | .mr(4) |
| 5344 | .nr(16) |
| 5345 | .kr(4) |
| 5346 | .sr(1) |
| 5347 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5348 | .n(n) |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5349 | .k(k) |
| 5350 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5351 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5352 | } |
| 5353 | } |
| 5354 | } |
| 5355 | |
| 5356 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_strided_a) { |
| 5357 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5358 | for (uint32_t n = 17; n < 32; n++) { |
| 5359 | for (size_t k = 1; k <= 80; k += 17) { |
| 5360 | GemmMicrokernelTester() |
| 5361 | .mr(4) |
| 5362 | .nr(16) |
| 5363 | .kr(4) |
| 5364 | .sr(1) |
| 5365 | .m(4) |
| 5366 | .n(n) |
| 5367 | .k(k) |
| 5368 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5369 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5370 | } |
| 5371 | } |
| 5372 | } |
| 5373 | |
| 5374 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_subtile) { |
| 5375 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5376 | for (uint32_t n = 17; n < 32; n++) { |
| 5377 | for (size_t k = 1; k <= 80; k += 17) { |
| 5378 | for (uint32_t m = 1; m <= 4; m++) { |
| 5379 | GemmMicrokernelTester() |
| 5380 | .mr(4) |
| 5381 | .nr(16) |
| 5382 | .kr(4) |
| 5383 | .sr(1) |
| 5384 | .m(m) |
| 5385 | .n(n) |
| 5386 | .k(k) |
| 5387 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5388 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5389 | } |
| 5390 | } |
| 5391 | } |
| 5392 | } |
| 5393 | |
| 5394 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16) { |
| 5395 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5396 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5397 | for (size_t k = 1; k <= 80; k += 17) { |
| 5398 | GemmMicrokernelTester() |
| 5399 | .mr(4) |
| 5400 | .nr(16) |
| 5401 | .kr(4) |
| 5402 | .sr(1) |
| 5403 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5404 | .n(n) |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5405 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5406 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5407 | } |
| 5408 | } |
| 5409 | } |
| 5410 | |
| 5411 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_strided_cn) { |
| 5412 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5413 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5414 | for (size_t k = 1; k <= 80; k += 17) { |
| 5415 | GemmMicrokernelTester() |
| 5416 | .mr(4) |
| 5417 | .nr(16) |
| 5418 | .kr(4) |
| 5419 | .sr(1) |
| 5420 | .m(4) |
| 5421 | .n(n) |
| 5422 | .k(k) |
| 5423 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5424 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5425 | } |
| 5426 | } |
| 5427 | } |
| 5428 | |
| 5429 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_strided_a) { |
| 5430 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5431 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5432 | for (size_t k = 1; k <= 80; k += 17) { |
| 5433 | GemmMicrokernelTester() |
| 5434 | .mr(4) |
| 5435 | .nr(16) |
| 5436 | .kr(4) |
| 5437 | .sr(1) |
| 5438 | .m(4) |
| 5439 | .n(n) |
| 5440 | .k(k) |
| 5441 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5442 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5443 | } |
| 5444 | } |
| 5445 | } |
| 5446 | |
| 5447 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_subtile) { |
| 5448 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5449 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5450 | for (size_t k = 1; k <= 80; k += 17) { |
| 5451 | for (uint32_t m = 1; m <= 4; m++) { |
| 5452 | GemmMicrokernelTester() |
| 5453 | .mr(4) |
| 5454 | .nr(16) |
| 5455 | .kr(4) |
| 5456 | .sr(1) |
| 5457 | .m(m) |
| 5458 | .n(n) |
| 5459 | .k(k) |
| 5460 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5461 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5462 | } |
| 5463 | } |
| 5464 | } |
| 5465 | } |
| 5466 | |
| 5467 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cm_subtile) { |
| 5468 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5469 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5470 | for (uint32_t n = 1; n <= 16; n++) { |
| 5471 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5472 | GemmMicrokernelTester() |
| 5473 | .mr(4) |
| 5474 | .nr(16) |
| 5475 | .kr(4) |
| 5476 | .sr(1) |
| 5477 | .m(m) |
| 5478 | .n(n) |
| 5479 | .k(k) |
| 5480 | .cm_stride(19) |
| 5481 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5482 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5483 | } |
| 5484 | } |
| 5485 | } |
| 5486 | } |
| 5487 | |
| 5488 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, qmin) { |
| 5489 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5490 | GemmMicrokernelTester() |
| 5491 | .mr(4) |
| 5492 | .nr(16) |
| 5493 | .kr(4) |
| 5494 | .sr(1) |
| 5495 | .m(4) |
| 5496 | .n(16) |
| 5497 | .k(16) |
| 5498 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5499 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5500 | } |
| 5501 | |
| 5502 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, qmax) { |
| 5503 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5504 | GemmMicrokernelTester() |
| 5505 | .mr(4) |
| 5506 | .nr(16) |
| 5507 | .kr(4) |
| 5508 | .sr(1) |
| 5509 | .m(4) |
| 5510 | .n(16) |
| 5511 | .k(16) |
| 5512 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5513 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5514 | } |
| 5515 | |
| 5516 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cm) { |
| 5517 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5518 | GemmMicrokernelTester() |
| 5519 | .mr(4) |
| 5520 | .nr(16) |
| 5521 | .kr(4) |
| 5522 | .sr(1) |
| 5523 | .m(4) |
| 5524 | .n(16) |
| 5525 | .k(16) |
| 5526 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5527 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5528 | } |
| 5529 | |
| 5530 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, no_a_zero_point) { |
| 5531 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5532 | for (size_t k = 1; k <= 80; k += 17) { |
| 5533 | GemmMicrokernelTester() |
| 5534 | .mr(4) |
| 5535 | .nr(16) |
| 5536 | .kr(4) |
| 5537 | .sr(1) |
| 5538 | .m(4) |
| 5539 | .n(16) |
| 5540 | .k(k) |
| 5541 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5542 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5543 | } |
| 5544 | } |
| 5545 | |
| 5546 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, no_b_zero_point) { |
| 5547 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5548 | for (size_t k = 1; k <= 80; k += 17) { |
| 5549 | GemmMicrokernelTester() |
| 5550 | .mr(4) |
| 5551 | .nr(16) |
| 5552 | .kr(4) |
| 5553 | .sr(1) |
| 5554 | .m(4) |
| 5555 | .n(16) |
| 5556 | .k(k) |
| 5557 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5558 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5559 | } |
| 5560 | } |
| 5561 | |
| 5562 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, no_zero_point) { |
| 5563 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5564 | for (size_t k = 1; k <= 80; k += 17) { |
| 5565 | GemmMicrokernelTester() |
| 5566 | .mr(4) |
| 5567 | .nr(16) |
| 5568 | .kr(4) |
| 5569 | .sr(1) |
| 5570 | .m(4) |
| 5571 | .n(16) |
| 5572 | .k(k) |
| 5573 | .a_zero_point(0) |
| 5574 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5575 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5576 | } |
| 5577 | } |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 5578 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | a49e41f | 2021-08-31 20:30:24 -0700 | [diff] [blame] | 5579 | |
| 5580 | |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 5581 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5582 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16) { |
| 5583 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5584 | GemmMicrokernelTester() |
| 5585 | .mr(4) |
| 5586 | .nr(16) |
| 5587 | .kr(4) |
| 5588 | .sr(1) |
| 5589 | .m(4) |
| 5590 | .n(16) |
| 5591 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5592 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5593 | } |
| 5594 | |
| 5595 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, strided_cn) { |
| 5596 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5597 | GemmMicrokernelTester() |
| 5598 | .mr(4) |
| 5599 | .nr(16) |
| 5600 | .kr(4) |
| 5601 | .sr(1) |
| 5602 | .m(4) |
| 5603 | .n(16) |
| 5604 | .k(16) |
| 5605 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5606 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5607 | } |
| 5608 | |
| 5609 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16_strided_a) { |
| 5610 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5611 | GemmMicrokernelTester() |
| 5612 | .mr(4) |
| 5613 | .nr(16) |
| 5614 | .kr(4) |
| 5615 | .sr(1) |
| 5616 | .m(4) |
| 5617 | .n(16) |
| 5618 | .k(16) |
| 5619 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5620 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5621 | } |
| 5622 | |
| 5623 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile) { |
| 5624 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5625 | for (uint32_t n = 1; n <= 16; n++) { |
| 5626 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5627 | GemmMicrokernelTester() |
| 5628 | .mr(4) |
| 5629 | .nr(16) |
| 5630 | .kr(4) |
| 5631 | .sr(1) |
| 5632 | .m(m) |
| 5633 | .n(n) |
| 5634 | .k(16) |
| 5635 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5636 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5637 | } |
| 5638 | } |
| 5639 | } |
| 5640 | |
| 5641 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile_m) { |
| 5642 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5643 | for (uint32_t m = 1; m <= 4; m++) { |
| 5644 | GemmMicrokernelTester() |
| 5645 | .mr(4) |
| 5646 | .nr(16) |
| 5647 | .kr(4) |
| 5648 | .sr(1) |
| 5649 | .m(m) |
| 5650 | .n(16) |
| 5651 | .k(16) |
| 5652 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5653 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5654 | } |
| 5655 | } |
| 5656 | |
| 5657 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile_n) { |
| 5658 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5659 | for (uint32_t n = 1; n <= 16; n++) { |
| 5660 | GemmMicrokernelTester() |
| 5661 | .mr(4) |
| 5662 | .nr(16) |
| 5663 | .kr(4) |
| 5664 | .sr(1) |
| 5665 | .m(4) |
| 5666 | .n(n) |
| 5667 | .k(16) |
| 5668 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5669 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5670 | } |
| 5671 | } |
| 5672 | |
| 5673 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_lt_16) { |
| 5674 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5675 | for (size_t k = 1; k < 16; k++) { |
| 5676 | GemmMicrokernelTester() |
| 5677 | .mr(4) |
| 5678 | .nr(16) |
| 5679 | .kr(4) |
| 5680 | .sr(1) |
| 5681 | .m(4) |
| 5682 | .n(16) |
| 5683 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5684 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5685 | } |
| 5686 | } |
| 5687 | |
| 5688 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_lt_16_strided_a) { |
| 5689 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5690 | for (size_t k = 1; k < 16; k++) { |
| 5691 | GemmMicrokernelTester() |
| 5692 | .mr(4) |
| 5693 | .nr(16) |
| 5694 | .kr(4) |
| 5695 | .sr(1) |
| 5696 | .m(4) |
| 5697 | .n(16) |
| 5698 | .k(k) |
| 5699 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5700 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5701 | } |
| 5702 | } |
| 5703 | |
| 5704 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_lt_16_subtile) { |
| 5705 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5706 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5707 | for (uint32_t n = 1; n <= 16; n++) { |
| 5708 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5709 | GemmMicrokernelTester() |
| 5710 | .mr(4) |
| 5711 | .nr(16) |
| 5712 | .kr(4) |
| 5713 | .sr(1) |
| 5714 | .m(m) |
| 5715 | .n(n) |
| 5716 | .k(k) |
| 5717 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5718 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5719 | } |
| 5720 | } |
| 5721 | } |
| 5722 | } |
| 5723 | |
| 5724 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_gt_16) { |
| 5725 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5726 | for (size_t k = 17; k < 32; k++) { |
| 5727 | GemmMicrokernelTester() |
| 5728 | .mr(4) |
| 5729 | .nr(16) |
| 5730 | .kr(4) |
| 5731 | .sr(1) |
| 5732 | .m(4) |
| 5733 | .n(16) |
| 5734 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5735 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5736 | } |
| 5737 | } |
| 5738 | |
| 5739 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_gt_16_strided_a) { |
| 5740 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5741 | for (size_t k = 17; k < 32; k++) { |
| 5742 | GemmMicrokernelTester() |
| 5743 | .mr(4) |
| 5744 | .nr(16) |
| 5745 | .kr(4) |
| 5746 | .sr(1) |
| 5747 | .m(4) |
| 5748 | .n(16) |
| 5749 | .k(k) |
| 5750 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5751 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5752 | } |
| 5753 | } |
| 5754 | |
| 5755 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_gt_16_subtile) { |
| 5756 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5757 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5758 | for (uint32_t n = 1; n <= 16; n++) { |
| 5759 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5760 | GemmMicrokernelTester() |
| 5761 | .mr(4) |
| 5762 | .nr(16) |
| 5763 | .kr(4) |
| 5764 | .sr(1) |
| 5765 | .m(m) |
| 5766 | .n(n) |
| 5767 | .k(k) |
| 5768 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5769 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5770 | } |
| 5771 | } |
| 5772 | } |
| 5773 | } |
| 5774 | |
| 5775 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_div_16) { |
| 5776 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5777 | for (size_t k = 32; k <= 160; k += 16) { |
| 5778 | GemmMicrokernelTester() |
| 5779 | .mr(4) |
| 5780 | .nr(16) |
| 5781 | .kr(4) |
| 5782 | .sr(1) |
| 5783 | .m(4) |
| 5784 | .n(16) |
| 5785 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5786 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5787 | } |
| 5788 | } |
| 5789 | |
| 5790 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_div_16_strided_a) { |
| 5791 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5792 | for (size_t k = 32; k <= 160; k += 16) { |
| 5793 | GemmMicrokernelTester() |
| 5794 | .mr(4) |
| 5795 | .nr(16) |
| 5796 | .kr(4) |
| 5797 | .sr(1) |
| 5798 | .m(4) |
| 5799 | .n(16) |
| 5800 | .k(k) |
| 5801 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5802 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5803 | } |
| 5804 | } |
| 5805 | |
| 5806 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_div_16_subtile) { |
| 5807 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5808 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5809 | for (uint32_t n = 1; n <= 16; n++) { |
| 5810 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5811 | GemmMicrokernelTester() |
| 5812 | .mr(4) |
| 5813 | .nr(16) |
| 5814 | .kr(4) |
| 5815 | .sr(1) |
| 5816 | .m(m) |
| 5817 | .n(n) |
| 5818 | .k(k) |
| 5819 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5820 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5821 | } |
| 5822 | } |
| 5823 | } |
| 5824 | } |
| 5825 | |
| 5826 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16) { |
| 5827 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5828 | for (uint32_t n = 17; n < 32; n++) { |
| 5829 | for (size_t k = 1; k <= 80; k += 17) { |
| 5830 | GemmMicrokernelTester() |
| 5831 | .mr(4) |
| 5832 | .nr(16) |
| 5833 | .kr(4) |
| 5834 | .sr(1) |
| 5835 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5836 | .n(n) |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5837 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5838 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5839 | } |
| 5840 | } |
| 5841 | } |
| 5842 | |
| 5843 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16_strided_cn) { |
| 5844 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5845 | for (uint32_t n = 17; n < 32; n++) { |
| 5846 | for (size_t k = 1; k <= 80; k += 17) { |
| 5847 | GemmMicrokernelTester() |
| 5848 | .mr(4) |
| 5849 | .nr(16) |
| 5850 | .kr(4) |
| 5851 | .sr(1) |
| 5852 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5853 | .n(n) |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5854 | .k(k) |
| 5855 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5856 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5857 | } |
| 5858 | } |
| 5859 | } |
| 5860 | |
| 5861 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16_strided_a) { |
| 5862 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5863 | for (uint32_t n = 17; n < 32; n++) { |
| 5864 | for (size_t k = 1; k <= 80; k += 17) { |
| 5865 | GemmMicrokernelTester() |
| 5866 | .mr(4) |
| 5867 | .nr(16) |
| 5868 | .kr(4) |
| 5869 | .sr(1) |
| 5870 | .m(4) |
| 5871 | .n(n) |
| 5872 | .k(k) |
| 5873 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5874 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5875 | } |
| 5876 | } |
| 5877 | } |
| 5878 | |
| 5879 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16_subtile) { |
| 5880 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5881 | for (uint32_t n = 17; n < 32; n++) { |
| 5882 | for (size_t k = 1; k <= 80; k += 17) { |
| 5883 | for (uint32_t m = 1; m <= 4; m++) { |
| 5884 | GemmMicrokernelTester() |
| 5885 | .mr(4) |
| 5886 | .nr(16) |
| 5887 | .kr(4) |
| 5888 | .sr(1) |
| 5889 | .m(m) |
| 5890 | .n(n) |
| 5891 | .k(k) |
| 5892 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5893 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5894 | } |
| 5895 | } |
| 5896 | } |
| 5897 | } |
| 5898 | |
| 5899 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16) { |
| 5900 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5901 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5902 | for (size_t k = 1; k <= 80; k += 17) { |
| 5903 | GemmMicrokernelTester() |
| 5904 | .mr(4) |
| 5905 | .nr(16) |
| 5906 | .kr(4) |
| 5907 | .sr(1) |
| 5908 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5909 | .n(n) |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5910 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5911 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5912 | } |
| 5913 | } |
| 5914 | } |
| 5915 | |
| 5916 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16_strided_cn) { |
| 5917 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5918 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5919 | for (size_t k = 1; k <= 80; k += 17) { |
| 5920 | GemmMicrokernelTester() |
| 5921 | .mr(4) |
| 5922 | .nr(16) |
| 5923 | .kr(4) |
| 5924 | .sr(1) |
| 5925 | .m(4) |
| 5926 | .n(n) |
| 5927 | .k(k) |
| 5928 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5929 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5930 | } |
| 5931 | } |
| 5932 | } |
| 5933 | |
| 5934 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16_strided_a) { |
| 5935 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5936 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5937 | for (size_t k = 1; k <= 80; k += 17) { |
| 5938 | GemmMicrokernelTester() |
| 5939 | .mr(4) |
| 5940 | .nr(16) |
| 5941 | .kr(4) |
| 5942 | .sr(1) |
| 5943 | .m(4) |
| 5944 | .n(n) |
| 5945 | .k(k) |
| 5946 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5947 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5948 | } |
| 5949 | } |
| 5950 | } |
| 5951 | |
| 5952 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16_subtile) { |
| 5953 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5954 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5955 | for (size_t k = 1; k <= 80; k += 17) { |
| 5956 | for (uint32_t m = 1; m <= 4; m++) { |
| 5957 | GemmMicrokernelTester() |
| 5958 | .mr(4) |
| 5959 | .nr(16) |
| 5960 | .kr(4) |
| 5961 | .sr(1) |
| 5962 | .m(m) |
| 5963 | .n(n) |
| 5964 | .k(k) |
| 5965 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5966 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5967 | } |
| 5968 | } |
| 5969 | } |
| 5970 | } |
| 5971 | |
| 5972 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, strided_cm_subtile) { |
| 5973 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5974 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5975 | for (uint32_t n = 1; n <= 16; n++) { |
| 5976 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5977 | GemmMicrokernelTester() |
| 5978 | .mr(4) |
| 5979 | .nr(16) |
| 5980 | .kr(4) |
| 5981 | .sr(1) |
| 5982 | .m(m) |
| 5983 | .n(n) |
| 5984 | .k(k) |
| 5985 | .cm_stride(19) |
| 5986 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5987 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 5988 | } |
| 5989 | } |
| 5990 | } |
| 5991 | } |
| 5992 | |
| 5993 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, qmin) { |
| 5994 | TEST_REQUIRES_ARM_NEON_DOT; |
| 5995 | GemmMicrokernelTester() |
| 5996 | .mr(4) |
| 5997 | .nr(16) |
| 5998 | .kr(4) |
| 5999 | .sr(1) |
| 6000 | .m(4) |
| 6001 | .n(16) |
| 6002 | .k(16) |
| 6003 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6004 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 6005 | } |
| 6006 | |
| 6007 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, qmax) { |
| 6008 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6009 | GemmMicrokernelTester() |
| 6010 | .mr(4) |
| 6011 | .nr(16) |
| 6012 | .kr(4) |
| 6013 | .sr(1) |
| 6014 | .m(4) |
| 6015 | .n(16) |
| 6016 | .k(16) |
| 6017 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6018 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 6019 | } |
| 6020 | |
| 6021 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, strided_cm) { |
| 6022 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6023 | GemmMicrokernelTester() |
| 6024 | .mr(4) |
| 6025 | .nr(16) |
| 6026 | .kr(4) |
| 6027 | .sr(1) |
| 6028 | .m(4) |
| 6029 | .n(16) |
| 6030 | .k(16) |
| 6031 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6032 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 6033 | } |
| 6034 | |
| 6035 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, no_a_zero_point) { |
| 6036 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6037 | for (size_t k = 1; k <= 80; k += 17) { |
| 6038 | GemmMicrokernelTester() |
| 6039 | .mr(4) |
| 6040 | .nr(16) |
| 6041 | .kr(4) |
| 6042 | .sr(1) |
| 6043 | .m(4) |
| 6044 | .n(16) |
| 6045 | .k(k) |
| 6046 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6047 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 6048 | } |
| 6049 | } |
| 6050 | |
| 6051 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, no_b_zero_point) { |
| 6052 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6053 | for (size_t k = 1; k <= 80; k += 17) { |
| 6054 | GemmMicrokernelTester() |
| 6055 | .mr(4) |
| 6056 | .nr(16) |
| 6057 | .kr(4) |
| 6058 | .sr(1) |
| 6059 | .m(4) |
| 6060 | .n(16) |
| 6061 | .k(k) |
| 6062 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6063 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 6064 | } |
| 6065 | } |
| 6066 | |
| 6067 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, no_zero_point) { |
| 6068 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6069 | for (size_t k = 1; k <= 80; k += 17) { |
| 6070 | GemmMicrokernelTester() |
| 6071 | .mr(4) |
| 6072 | .nr(16) |
| 6073 | .kr(4) |
| 6074 | .sr(1) |
| 6075 | .m(4) |
| 6076 | .n(16) |
| 6077 | .k(k) |
| 6078 | .a_zero_point(0) |
| 6079 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6080 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 6081 | } |
| 6082 | } |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 6083 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | 4066898 | 2021-08-24 11:12:04 -0700 | [diff] [blame] | 6084 | |
| 6085 | |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 6086 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6087 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_eq_16) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6088 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6089 | GemmMicrokernelTester() |
| 6090 | .mr(4) |
| 6091 | .nr(8) |
| 6092 | .kr(4) |
| 6093 | .sr(1) |
| 6094 | .m(4) |
| 6095 | .n(8) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6096 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6097 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6098 | } |
| 6099 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6100 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, strided_cn) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6101 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6102 | GemmMicrokernelTester() |
| 6103 | .mr(4) |
| 6104 | .nr(8) |
| 6105 | .kr(4) |
| 6106 | .sr(1) |
| 6107 | .m(4) |
| 6108 | .n(8) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6109 | .k(16) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6110 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6111 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6112 | } |
| 6113 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6114 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_eq_16_strided_a) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6115 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6116 | GemmMicrokernelTester() |
| 6117 | .mr(4) |
| 6118 | .nr(8) |
| 6119 | .kr(4) |
| 6120 | .sr(1) |
| 6121 | .m(4) |
| 6122 | .n(8) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6123 | .k(16) |
| 6124 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6125 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6126 | } |
| 6127 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6128 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6129 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6130 | for (uint32_t n = 1; n <= 8; n++) { |
| 6131 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6132 | GemmMicrokernelTester() |
| 6133 | .mr(4) |
| 6134 | .nr(8) |
| 6135 | .kr(4) |
| 6136 | .sr(1) |
| 6137 | .m(m) |
| 6138 | .n(n) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6139 | .k(16) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6140 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6141 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6142 | } |
| 6143 | } |
| 6144 | } |
| 6145 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6146 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile_m) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6147 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6148 | for (uint32_t m = 1; m <= 4; m++) { |
| 6149 | GemmMicrokernelTester() |
| 6150 | .mr(4) |
| 6151 | .nr(8) |
| 6152 | .kr(4) |
| 6153 | .sr(1) |
| 6154 | .m(m) |
| 6155 | .n(8) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6156 | .k(16) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6157 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6158 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6159 | } |
| 6160 | } |
| 6161 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6162 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile_n) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6163 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6164 | for (uint32_t n = 1; n <= 8; n++) { |
| 6165 | GemmMicrokernelTester() |
| 6166 | .mr(4) |
| 6167 | .nr(8) |
| 6168 | .kr(4) |
| 6169 | .sr(1) |
| 6170 | .m(4) |
| 6171 | .n(n) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6172 | .k(16) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6173 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6174 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6175 | } |
| 6176 | } |
| 6177 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6178 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_lt_16) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6179 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6180 | for (size_t k = 1; k < 16; k++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6181 | GemmMicrokernelTester() |
| 6182 | .mr(4) |
| 6183 | .nr(8) |
| 6184 | .kr(4) |
| 6185 | .sr(1) |
| 6186 | .m(4) |
| 6187 | .n(8) |
| 6188 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6189 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6190 | } |
| 6191 | } |
| 6192 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6193 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_lt_16_strided_a) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6194 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6195 | for (size_t k = 1; k < 16; k++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6196 | GemmMicrokernelTester() |
| 6197 | .mr(4) |
| 6198 | .nr(8) |
| 6199 | .kr(4) |
| 6200 | .sr(1) |
| 6201 | .m(4) |
| 6202 | .n(8) |
| 6203 | .k(k) |
| 6204 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6205 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6206 | } |
| 6207 | } |
| 6208 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6209 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_lt_16_subtile) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6210 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6211 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6212 | for (uint32_t n = 1; n <= 8; n++) { |
| 6213 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6214 | GemmMicrokernelTester() |
| 6215 | .mr(4) |
| 6216 | .nr(8) |
| 6217 | .kr(4) |
| 6218 | .sr(1) |
| 6219 | .m(m) |
| 6220 | .n(n) |
| 6221 | .k(k) |
| 6222 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6223 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6224 | } |
| 6225 | } |
| 6226 | } |
| 6227 | } |
| 6228 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6229 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_gt_16) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6230 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6231 | for (size_t k = 17; k < 32; k++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6232 | GemmMicrokernelTester() |
| 6233 | .mr(4) |
| 6234 | .nr(8) |
| 6235 | .kr(4) |
| 6236 | .sr(1) |
| 6237 | .m(4) |
| 6238 | .n(8) |
| 6239 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6240 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6241 | } |
| 6242 | } |
| 6243 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6244 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_gt_16_strided_a) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6245 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6246 | for (size_t k = 17; k < 32; k++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6247 | GemmMicrokernelTester() |
| 6248 | .mr(4) |
| 6249 | .nr(8) |
| 6250 | .kr(4) |
| 6251 | .sr(1) |
| 6252 | .m(4) |
| 6253 | .n(8) |
| 6254 | .k(k) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6255 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6256 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6257 | } |
| 6258 | } |
| 6259 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6260 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_gt_16_subtile) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6261 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6262 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6263 | for (uint32_t n = 1; n <= 8; n++) { |
| 6264 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6265 | GemmMicrokernelTester() |
| 6266 | .mr(4) |
| 6267 | .nr(8) |
| 6268 | .kr(4) |
| 6269 | .sr(1) |
| 6270 | .m(m) |
| 6271 | .n(n) |
| 6272 | .k(k) |
| 6273 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6274 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6275 | } |
| 6276 | } |
| 6277 | } |
| 6278 | } |
| 6279 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6280 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_div_16) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6281 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6282 | for (size_t k = 32; k <= 160; k += 16) { |
| 6283 | GemmMicrokernelTester() |
| 6284 | .mr(4) |
| 6285 | .nr(8) |
| 6286 | .kr(4) |
| 6287 | .sr(1) |
| 6288 | .m(4) |
| 6289 | .n(8) |
| 6290 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6291 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6292 | } |
| 6293 | } |
| 6294 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6295 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_div_16_strided_a) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6296 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6297 | for (size_t k = 32; k <= 160; k += 16) { |
| 6298 | GemmMicrokernelTester() |
| 6299 | .mr(4) |
| 6300 | .nr(8) |
| 6301 | .kr(4) |
| 6302 | .sr(1) |
| 6303 | .m(4) |
| 6304 | .n(8) |
| 6305 | .k(k) |
| 6306 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6307 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6308 | } |
| 6309 | } |
| 6310 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6311 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, k_div_16_subtile) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6312 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6313 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6314 | for (uint32_t n = 1; n <= 8; n++) { |
| 6315 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6316 | GemmMicrokernelTester() |
| 6317 | .mr(4) |
| 6318 | .nr(8) |
| 6319 | .kr(4) |
| 6320 | .sr(1) |
| 6321 | .m(m) |
| 6322 | .n(n) |
| 6323 | .k(k) |
| 6324 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6325 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6326 | } |
| 6327 | } |
| 6328 | } |
| 6329 | } |
| 6330 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6331 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_gt_8) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6332 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6333 | for (uint32_t n = 9; n < 16; n++) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6334 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6335 | GemmMicrokernelTester() |
| 6336 | .mr(4) |
| 6337 | .nr(8) |
| 6338 | .kr(4) |
| 6339 | .sr(1) |
| 6340 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6341 | .n(n) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6342 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6343 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6344 | } |
| 6345 | } |
| 6346 | } |
| 6347 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6348 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_gt_8_strided_cn) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6349 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6350 | for (uint32_t n = 9; n < 16; n++) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6351 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6352 | GemmMicrokernelTester() |
| 6353 | .mr(4) |
| 6354 | .nr(8) |
| 6355 | .kr(4) |
| 6356 | .sr(1) |
| 6357 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6358 | .n(n) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6359 | .k(k) |
| 6360 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6361 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6362 | } |
| 6363 | } |
| 6364 | } |
| 6365 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6366 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_gt_8_strided_a) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6367 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6368 | for (uint32_t n = 9; n < 16; n++) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6369 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6370 | GemmMicrokernelTester() |
| 6371 | .mr(4) |
| 6372 | .nr(8) |
| 6373 | .kr(4) |
| 6374 | .sr(1) |
| 6375 | .m(4) |
| 6376 | .n(n) |
| 6377 | .k(k) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6378 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6379 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6380 | } |
| 6381 | } |
| 6382 | } |
| 6383 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6384 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_gt_8_subtile) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6385 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6386 | for (uint32_t n = 9; n < 16; n++) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6387 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6388 | for (uint32_t m = 1; m <= 4; m++) { |
| 6389 | GemmMicrokernelTester() |
| 6390 | .mr(4) |
| 6391 | .nr(8) |
| 6392 | .kr(4) |
| 6393 | .sr(1) |
| 6394 | .m(m) |
| 6395 | .n(n) |
| 6396 | .k(k) |
| 6397 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6398 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6399 | } |
| 6400 | } |
| 6401 | } |
| 6402 | } |
| 6403 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6404 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_div_8) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6405 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6406 | for (uint32_t n = 16; n <= 24; n += 8) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6407 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6408 | GemmMicrokernelTester() |
| 6409 | .mr(4) |
| 6410 | .nr(8) |
| 6411 | .kr(4) |
| 6412 | .sr(1) |
| 6413 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6414 | .n(n) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6415 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6416 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6417 | } |
| 6418 | } |
| 6419 | } |
| 6420 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6421 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_div_8_strided_cn) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6422 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6423 | for (uint32_t n = 16; n <= 24; n += 8) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6424 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6425 | GemmMicrokernelTester() |
| 6426 | .mr(4) |
| 6427 | .nr(8) |
| 6428 | .kr(4) |
| 6429 | .sr(1) |
| 6430 | .m(4) |
| 6431 | .n(n) |
| 6432 | .k(k) |
| 6433 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6434 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6435 | } |
| 6436 | } |
| 6437 | } |
| 6438 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6439 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_div_8_strided_a) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6440 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6441 | for (uint32_t n = 16; n <= 24; n += 8) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6442 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6443 | GemmMicrokernelTester() |
| 6444 | .mr(4) |
| 6445 | .nr(8) |
| 6446 | .kr(4) |
| 6447 | .sr(1) |
| 6448 | .m(4) |
| 6449 | .n(n) |
| 6450 | .k(k) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6451 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6452 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6453 | } |
| 6454 | } |
| 6455 | } |
| 6456 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6457 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, n_div_8_subtile) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6458 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6459 | for (uint32_t n = 16; n <= 24; n += 8) { |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6460 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6461 | for (uint32_t m = 1; m <= 4; m++) { |
| 6462 | GemmMicrokernelTester() |
| 6463 | .mr(4) |
| 6464 | .nr(8) |
| 6465 | .kr(4) |
| 6466 | .sr(1) |
| 6467 | .m(m) |
| 6468 | .n(n) |
| 6469 | .k(k) |
| 6470 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6471 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6472 | } |
| 6473 | } |
| 6474 | } |
| 6475 | } |
| 6476 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6477 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, strided_cm_subtile) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6478 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6479 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6480 | for (uint32_t n = 1; n <= 8; n++) { |
| 6481 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6482 | GemmMicrokernelTester() |
| 6483 | .mr(4) |
| 6484 | .nr(8) |
| 6485 | .kr(4) |
| 6486 | .sr(1) |
| 6487 | .m(m) |
| 6488 | .n(n) |
| 6489 | .k(k) |
| 6490 | .cm_stride(11) |
| 6491 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6492 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6493 | } |
| 6494 | } |
| 6495 | } |
| 6496 | } |
| 6497 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6498 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, qmin) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6499 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6500 | GemmMicrokernelTester() |
| 6501 | .mr(4) |
| 6502 | .nr(8) |
| 6503 | .kr(4) |
| 6504 | .sr(1) |
| 6505 | .m(4) |
| 6506 | .n(8) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6507 | .k(16) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6508 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6509 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6510 | } |
| 6511 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6512 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, qmax) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6513 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6514 | GemmMicrokernelTester() |
| 6515 | .mr(4) |
| 6516 | .nr(8) |
| 6517 | .kr(4) |
| 6518 | .sr(1) |
| 6519 | .m(4) |
| 6520 | .n(8) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6521 | .k(16) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6522 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6523 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6524 | } |
| 6525 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6526 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, strided_cm) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6527 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6528 | GemmMicrokernelTester() |
| 6529 | .mr(4) |
| 6530 | .nr(8) |
| 6531 | .kr(4) |
| 6532 | .sr(1) |
| 6533 | .m(4) |
| 6534 | .n(8) |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6535 | .k(16) |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6536 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6537 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6538 | } |
| 6539 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6540 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, no_a_zero_point) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6541 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6542 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6543 | GemmMicrokernelTester() |
| 6544 | .mr(4) |
| 6545 | .nr(8) |
| 6546 | .kr(4) |
| 6547 | .sr(1) |
| 6548 | .m(4) |
| 6549 | .n(8) |
| 6550 | .k(k) |
| 6551 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6552 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6553 | } |
| 6554 | } |
| 6555 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6556 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, no_b_zero_point) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6557 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6558 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6559 | GemmMicrokernelTester() |
| 6560 | .mr(4) |
| 6561 | .nr(8) |
| 6562 | .kr(4) |
| 6563 | .sr(1) |
| 6564 | .m(4) |
| 6565 | .n(8) |
| 6566 | .k(k) |
| 6567 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6568 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6569 | } |
| 6570 | } |
| 6571 | |
Frank Barchard | df8e604 | 2021-09-03 13:56:29 -0700 | [diff] [blame] | 6572 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__AARCH64_NEONDOT_LD128, no_zero_point) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6573 | TEST_REQUIRES_ARM_NEON_DOT; |
Frank Barchard | bd5b027 | 2021-09-03 10:16:21 -0700 | [diff] [blame] | 6574 | for (size_t k = 1; k <= 80; k += 17) { |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6575 | GemmMicrokernelTester() |
| 6576 | .mr(4) |
| 6577 | .nr(8) |
| 6578 | .kr(4) |
| 6579 | .sr(1) |
| 6580 | .m(4) |
| 6581 | .n(8) |
| 6582 | .k(k) |
| 6583 | .a_zero_point(0) |
| 6584 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6585 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__aarch64_neondot_ld128, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6586 | } |
| 6587 | } |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 6588 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | 0049e89 | 2021-08-22 09:37:21 -0700 | [diff] [blame] | 6589 | |
| 6590 | |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 6591 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6592 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8) { |
| 6593 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6594 | GemmMicrokernelTester() |
| 6595 | .mr(2) |
| 6596 | .nr(8) |
| 6597 | .kr(4) |
| 6598 | .sr(1) |
| 6599 | .m(2) |
| 6600 | .n(8) |
| 6601 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6602 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6603 | } |
| 6604 | |
| 6605 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cn) { |
| 6606 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6607 | GemmMicrokernelTester() |
| 6608 | .mr(2) |
| 6609 | .nr(8) |
| 6610 | .kr(4) |
| 6611 | .sr(1) |
| 6612 | .m(2) |
| 6613 | .n(8) |
| 6614 | .k(8) |
| 6615 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6616 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6617 | } |
| 6618 | |
| 6619 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_strided_a) { |
| 6620 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6621 | GemmMicrokernelTester() |
| 6622 | .mr(2) |
| 6623 | .nr(8) |
| 6624 | .kr(4) |
| 6625 | .sr(1) |
| 6626 | .m(2) |
| 6627 | .n(8) |
| 6628 | .k(8) |
| 6629 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6630 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6631 | } |
| 6632 | |
| 6633 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile) { |
| 6634 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6635 | for (uint32_t n = 1; n <= 8; n++) { |
| 6636 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6637 | GemmMicrokernelTester() |
| 6638 | .mr(2) |
| 6639 | .nr(8) |
| 6640 | .kr(4) |
| 6641 | .sr(1) |
| 6642 | .m(m) |
| 6643 | .n(n) |
| 6644 | .k(8) |
| 6645 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6646 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6647 | } |
| 6648 | } |
| 6649 | } |
| 6650 | |
| 6651 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile_m) { |
| 6652 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6653 | for (uint32_t m = 1; m <= 2; m++) { |
| 6654 | GemmMicrokernelTester() |
| 6655 | .mr(2) |
| 6656 | .nr(8) |
| 6657 | .kr(4) |
| 6658 | .sr(1) |
| 6659 | .m(m) |
| 6660 | .n(8) |
| 6661 | .k(8) |
| 6662 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6663 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6664 | } |
| 6665 | } |
| 6666 | |
| 6667 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_eq_8_subtile_n) { |
| 6668 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6669 | for (uint32_t n = 1; n <= 8; n++) { |
| 6670 | GemmMicrokernelTester() |
| 6671 | .mr(2) |
| 6672 | .nr(8) |
| 6673 | .kr(4) |
| 6674 | .sr(1) |
| 6675 | .m(2) |
| 6676 | .n(n) |
| 6677 | .k(8) |
| 6678 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6679 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6680 | } |
| 6681 | } |
| 6682 | |
| 6683 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8) { |
| 6684 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6685 | for (size_t k = 1; k < 8; k++) { |
| 6686 | GemmMicrokernelTester() |
| 6687 | .mr(2) |
| 6688 | .nr(8) |
| 6689 | .kr(4) |
| 6690 | .sr(1) |
| 6691 | .m(2) |
| 6692 | .n(8) |
| 6693 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6694 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6695 | } |
| 6696 | } |
| 6697 | |
| 6698 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8_strided_a) { |
| 6699 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6700 | for (size_t k = 1; k < 8; k++) { |
| 6701 | GemmMicrokernelTester() |
| 6702 | .mr(2) |
| 6703 | .nr(8) |
| 6704 | .kr(4) |
| 6705 | .sr(1) |
| 6706 | .m(2) |
| 6707 | .n(8) |
| 6708 | .k(k) |
| 6709 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6710 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6711 | } |
| 6712 | } |
| 6713 | |
| 6714 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_lt_8_subtile) { |
| 6715 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6716 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6717 | for (uint32_t n = 1; n <= 8; n++) { |
| 6718 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6719 | GemmMicrokernelTester() |
| 6720 | .mr(2) |
| 6721 | .nr(8) |
| 6722 | .kr(4) |
| 6723 | .sr(1) |
| 6724 | .m(m) |
| 6725 | .n(n) |
| 6726 | .k(k) |
| 6727 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6728 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6729 | } |
| 6730 | } |
| 6731 | } |
| 6732 | } |
| 6733 | |
| 6734 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8) { |
| 6735 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6736 | for (size_t k = 9; k < 16; k++) { |
| 6737 | GemmMicrokernelTester() |
| 6738 | .mr(2) |
| 6739 | .nr(8) |
| 6740 | .kr(4) |
| 6741 | .sr(1) |
| 6742 | .m(2) |
| 6743 | .n(8) |
| 6744 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6745 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6746 | } |
| 6747 | } |
| 6748 | |
| 6749 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8_strided_a) { |
| 6750 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6751 | for (size_t k = 9; k < 16; k++) { |
| 6752 | GemmMicrokernelTester() |
| 6753 | .mr(2) |
| 6754 | .nr(8) |
| 6755 | .kr(4) |
| 6756 | .sr(1) |
| 6757 | .m(2) |
| 6758 | .n(8) |
| 6759 | .k(k) |
| 6760 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6761 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6762 | } |
| 6763 | } |
| 6764 | |
| 6765 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_gt_8_subtile) { |
| 6766 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6767 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6768 | for (uint32_t n = 1; n <= 8; n++) { |
| 6769 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6770 | GemmMicrokernelTester() |
| 6771 | .mr(2) |
| 6772 | .nr(8) |
| 6773 | .kr(4) |
| 6774 | .sr(1) |
| 6775 | .m(m) |
| 6776 | .n(n) |
| 6777 | .k(k) |
| 6778 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6779 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6780 | } |
| 6781 | } |
| 6782 | } |
| 6783 | } |
| 6784 | |
| 6785 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8) { |
| 6786 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6787 | for (size_t k = 16; k <= 80; k += 8) { |
| 6788 | GemmMicrokernelTester() |
| 6789 | .mr(2) |
| 6790 | .nr(8) |
| 6791 | .kr(4) |
| 6792 | .sr(1) |
| 6793 | .m(2) |
| 6794 | .n(8) |
| 6795 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6796 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6797 | } |
| 6798 | } |
| 6799 | |
| 6800 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8_strided_a) { |
| 6801 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6802 | for (size_t k = 16; k <= 80; k += 8) { |
| 6803 | GemmMicrokernelTester() |
| 6804 | .mr(2) |
| 6805 | .nr(8) |
| 6806 | .kr(4) |
| 6807 | .sr(1) |
| 6808 | .m(2) |
| 6809 | .n(8) |
| 6810 | .k(k) |
| 6811 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6812 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6813 | } |
| 6814 | } |
| 6815 | |
| 6816 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, k_div_8_subtile) { |
| 6817 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6818 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6819 | for (uint32_t n = 1; n <= 8; n++) { |
| 6820 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6821 | GemmMicrokernelTester() |
| 6822 | .mr(2) |
| 6823 | .nr(8) |
| 6824 | .kr(4) |
| 6825 | .sr(1) |
| 6826 | .m(m) |
| 6827 | .n(n) |
| 6828 | .k(k) |
| 6829 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6830 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6831 | } |
| 6832 | } |
| 6833 | } |
| 6834 | } |
| 6835 | |
| 6836 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8) { |
| 6837 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6838 | for (uint32_t n = 9; n < 16; n++) { |
| 6839 | for (size_t k = 1; k <= 40; k += 9) { |
| 6840 | GemmMicrokernelTester() |
| 6841 | .mr(2) |
| 6842 | .nr(8) |
| 6843 | .kr(4) |
| 6844 | .sr(1) |
| 6845 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6846 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6847 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6848 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6849 | } |
| 6850 | } |
| 6851 | } |
| 6852 | |
| 6853 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_strided_cn) { |
| 6854 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6855 | for (uint32_t n = 9; n < 16; n++) { |
| 6856 | for (size_t k = 1; k <= 40; k += 9) { |
| 6857 | GemmMicrokernelTester() |
| 6858 | .mr(2) |
| 6859 | .nr(8) |
| 6860 | .kr(4) |
| 6861 | .sr(1) |
| 6862 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6863 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6864 | .k(k) |
| 6865 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6866 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6867 | } |
| 6868 | } |
| 6869 | } |
| 6870 | |
| 6871 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_strided_a) { |
| 6872 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6873 | for (uint32_t n = 9; n < 16; n++) { |
| 6874 | for (size_t k = 1; k <= 40; k += 9) { |
| 6875 | GemmMicrokernelTester() |
| 6876 | .mr(2) |
| 6877 | .nr(8) |
| 6878 | .kr(4) |
| 6879 | .sr(1) |
| 6880 | .m(2) |
| 6881 | .n(n) |
| 6882 | .k(k) |
| 6883 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6884 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6885 | } |
| 6886 | } |
| 6887 | } |
| 6888 | |
| 6889 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_gt_8_subtile) { |
| 6890 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6891 | for (uint32_t n = 9; n < 16; n++) { |
| 6892 | for (size_t k = 1; k <= 40; k += 9) { |
| 6893 | for (uint32_t m = 1; m <= 2; m++) { |
| 6894 | GemmMicrokernelTester() |
| 6895 | .mr(2) |
| 6896 | .nr(8) |
| 6897 | .kr(4) |
| 6898 | .sr(1) |
| 6899 | .m(m) |
| 6900 | .n(n) |
| 6901 | .k(k) |
| 6902 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6903 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6904 | } |
| 6905 | } |
| 6906 | } |
| 6907 | } |
| 6908 | |
| 6909 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8) { |
| 6910 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6911 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6912 | for (size_t k = 1; k <= 40; k += 9) { |
| 6913 | GemmMicrokernelTester() |
| 6914 | .mr(2) |
| 6915 | .nr(8) |
| 6916 | .kr(4) |
| 6917 | .sr(1) |
| 6918 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6919 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6920 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6921 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6922 | } |
| 6923 | } |
| 6924 | } |
| 6925 | |
| 6926 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_strided_cn) { |
| 6927 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6928 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6929 | for (size_t k = 1; k <= 40; k += 9) { |
| 6930 | GemmMicrokernelTester() |
| 6931 | .mr(2) |
| 6932 | .nr(8) |
| 6933 | .kr(4) |
| 6934 | .sr(1) |
| 6935 | .m(2) |
| 6936 | .n(n) |
| 6937 | .k(k) |
| 6938 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6939 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6940 | } |
| 6941 | } |
| 6942 | } |
| 6943 | |
| 6944 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_strided_a) { |
| 6945 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6946 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6947 | for (size_t k = 1; k <= 40; k += 9) { |
| 6948 | GemmMicrokernelTester() |
| 6949 | .mr(2) |
| 6950 | .nr(8) |
| 6951 | .kr(4) |
| 6952 | .sr(1) |
| 6953 | .m(2) |
| 6954 | .n(n) |
| 6955 | .k(k) |
| 6956 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6957 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6958 | } |
| 6959 | } |
| 6960 | } |
| 6961 | |
| 6962 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, n_div_8_subtile) { |
| 6963 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6964 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6965 | for (size_t k = 1; k <= 40; k += 9) { |
| 6966 | for (uint32_t m = 1; m <= 2; m++) { |
| 6967 | GemmMicrokernelTester() |
| 6968 | .mr(2) |
| 6969 | .nr(8) |
| 6970 | .kr(4) |
| 6971 | .sr(1) |
| 6972 | .m(m) |
| 6973 | .n(n) |
| 6974 | .k(k) |
| 6975 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6976 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6977 | } |
| 6978 | } |
| 6979 | } |
| 6980 | } |
| 6981 | |
| 6982 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cm_subtile) { |
| 6983 | TEST_REQUIRES_ARM_NEON_DOT; |
| 6984 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6985 | for (uint32_t n = 1; n <= 8; n++) { |
| 6986 | for (uint32_t m = 1; m <= 2; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6987 | GemmMicrokernelTester() |
| 6988 | .mr(2) |
| 6989 | .nr(8) |
| 6990 | .kr(4) |
| 6991 | .sr(1) |
| 6992 | .m(m) |
| 6993 | .n(n) |
| 6994 | .k(k) |
| 6995 | .cm_stride(11) |
| 6996 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6997 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 6998 | } |
| 6999 | } |
| 7000 | } |
| 7001 | } |
| 7002 | |
| 7003 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, qmin) { |
| 7004 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7005 | GemmMicrokernelTester() |
| 7006 | .mr(2) |
| 7007 | .nr(8) |
| 7008 | .kr(4) |
| 7009 | .sr(1) |
| 7010 | .m(2) |
| 7011 | .n(8) |
| 7012 | .k(8) |
| 7013 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7014 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7015 | } |
| 7016 | |
| 7017 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, qmax) { |
| 7018 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7019 | GemmMicrokernelTester() |
| 7020 | .mr(2) |
| 7021 | .nr(8) |
| 7022 | .kr(4) |
| 7023 | .sr(1) |
| 7024 | .m(2) |
| 7025 | .n(8) |
| 7026 | .k(8) |
| 7027 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7028 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7029 | } |
| 7030 | |
| 7031 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, strided_cm) { |
| 7032 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7033 | GemmMicrokernelTester() |
| 7034 | .mr(2) |
| 7035 | .nr(8) |
| 7036 | .kr(4) |
| 7037 | .sr(1) |
| 7038 | .m(2) |
| 7039 | .n(8) |
| 7040 | .k(8) |
| 7041 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7042 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7043 | } |
| 7044 | |
| 7045 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_a_zero_point) { |
| 7046 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7047 | for (size_t k = 1; k <= 40; k += 9) { |
| 7048 | GemmMicrokernelTester() |
| 7049 | .mr(2) |
| 7050 | .nr(8) |
| 7051 | .kr(4) |
| 7052 | .sr(1) |
| 7053 | .m(2) |
| 7054 | .n(8) |
| 7055 | .k(k) |
| 7056 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7057 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7058 | } |
| 7059 | } |
| 7060 | |
| 7061 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_b_zero_point) { |
| 7062 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7063 | for (size_t k = 1; k <= 40; k += 9) { |
| 7064 | GemmMicrokernelTester() |
| 7065 | .mr(2) |
| 7066 | .nr(8) |
| 7067 | .kr(4) |
| 7068 | .sr(1) |
| 7069 | .m(2) |
| 7070 | .n(8) |
| 7071 | .k(k) |
| 7072 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7073 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7074 | } |
| 7075 | } |
| 7076 | |
| 7077 | TEST(QU8_GEMM_MINMAX_RNDNU_2X8C4__NEONDOT, no_zero_point) { |
| 7078 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7079 | for (size_t k = 1; k <= 40; k += 9) { |
| 7080 | GemmMicrokernelTester() |
| 7081 | .mr(2) |
| 7082 | .nr(8) |
| 7083 | .kr(4) |
| 7084 | .sr(1) |
| 7085 | .m(2) |
| 7086 | .n(8) |
| 7087 | .k(k) |
| 7088 | .a_zero_point(0) |
| 7089 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7090 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_2x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7091 | } |
| 7092 | } |
| 7093 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 7094 | |
| 7095 | |
| 7096 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 7097 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8) { |
| 7098 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7099 | GemmMicrokernelTester() |
| 7100 | .mr(3) |
| 7101 | .nr(8) |
| 7102 | .kr(4) |
| 7103 | .sr(1) |
| 7104 | .m(3) |
| 7105 | .n(8) |
| 7106 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7107 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7108 | } |
| 7109 | |
| 7110 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cn) { |
| 7111 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7112 | GemmMicrokernelTester() |
| 7113 | .mr(3) |
| 7114 | .nr(8) |
| 7115 | .kr(4) |
| 7116 | .sr(1) |
| 7117 | .m(3) |
| 7118 | .n(8) |
| 7119 | .k(8) |
| 7120 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7121 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7122 | } |
| 7123 | |
| 7124 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_strided_a) { |
| 7125 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7126 | GemmMicrokernelTester() |
| 7127 | .mr(3) |
| 7128 | .nr(8) |
| 7129 | .kr(4) |
| 7130 | .sr(1) |
| 7131 | .m(3) |
| 7132 | .n(8) |
| 7133 | .k(8) |
| 7134 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7135 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7136 | } |
| 7137 | |
| 7138 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile) { |
| 7139 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7140 | for (uint32_t n = 1; n <= 8; n++) { |
| 7141 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7142 | GemmMicrokernelTester() |
| 7143 | .mr(3) |
| 7144 | .nr(8) |
| 7145 | .kr(4) |
| 7146 | .sr(1) |
| 7147 | .m(m) |
| 7148 | .n(n) |
| 7149 | .k(8) |
| 7150 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7151 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7152 | } |
| 7153 | } |
| 7154 | } |
| 7155 | |
| 7156 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile_m) { |
| 7157 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7158 | for (uint32_t m = 1; m <= 3; m++) { |
| 7159 | GemmMicrokernelTester() |
| 7160 | .mr(3) |
| 7161 | .nr(8) |
| 7162 | .kr(4) |
| 7163 | .sr(1) |
| 7164 | .m(m) |
| 7165 | .n(8) |
| 7166 | .k(8) |
| 7167 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7168 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7169 | } |
| 7170 | } |
| 7171 | |
| 7172 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_eq_8_subtile_n) { |
| 7173 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7174 | for (uint32_t n = 1; n <= 8; n++) { |
| 7175 | GemmMicrokernelTester() |
| 7176 | .mr(3) |
| 7177 | .nr(8) |
| 7178 | .kr(4) |
| 7179 | .sr(1) |
| 7180 | .m(3) |
| 7181 | .n(n) |
| 7182 | .k(8) |
| 7183 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7184 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7185 | } |
| 7186 | } |
| 7187 | |
| 7188 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8) { |
| 7189 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7190 | for (size_t k = 1; k < 8; k++) { |
| 7191 | GemmMicrokernelTester() |
| 7192 | .mr(3) |
| 7193 | .nr(8) |
| 7194 | .kr(4) |
| 7195 | .sr(1) |
| 7196 | .m(3) |
| 7197 | .n(8) |
| 7198 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7199 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7200 | } |
| 7201 | } |
| 7202 | |
| 7203 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8_strided_a) { |
| 7204 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7205 | for (size_t k = 1; k < 8; k++) { |
| 7206 | GemmMicrokernelTester() |
| 7207 | .mr(3) |
| 7208 | .nr(8) |
| 7209 | .kr(4) |
| 7210 | .sr(1) |
| 7211 | .m(3) |
| 7212 | .n(8) |
| 7213 | .k(k) |
| 7214 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7215 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7216 | } |
| 7217 | } |
| 7218 | |
| 7219 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_lt_8_subtile) { |
| 7220 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7221 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7222 | for (uint32_t n = 1; n <= 8; n++) { |
| 7223 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7224 | GemmMicrokernelTester() |
| 7225 | .mr(3) |
| 7226 | .nr(8) |
| 7227 | .kr(4) |
| 7228 | .sr(1) |
| 7229 | .m(m) |
| 7230 | .n(n) |
| 7231 | .k(k) |
| 7232 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7233 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7234 | } |
| 7235 | } |
| 7236 | } |
| 7237 | } |
| 7238 | |
| 7239 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8) { |
| 7240 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7241 | for (size_t k = 9; k < 16; k++) { |
| 7242 | GemmMicrokernelTester() |
| 7243 | .mr(3) |
| 7244 | .nr(8) |
| 7245 | .kr(4) |
| 7246 | .sr(1) |
| 7247 | .m(3) |
| 7248 | .n(8) |
| 7249 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7250 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7251 | } |
| 7252 | } |
| 7253 | |
| 7254 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8_strided_a) { |
| 7255 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7256 | for (size_t k = 9; k < 16; k++) { |
| 7257 | GemmMicrokernelTester() |
| 7258 | .mr(3) |
| 7259 | .nr(8) |
| 7260 | .kr(4) |
| 7261 | .sr(1) |
| 7262 | .m(3) |
| 7263 | .n(8) |
| 7264 | .k(k) |
| 7265 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7266 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7267 | } |
| 7268 | } |
| 7269 | |
| 7270 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_gt_8_subtile) { |
| 7271 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7272 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7273 | for (uint32_t n = 1; n <= 8; n++) { |
| 7274 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7275 | GemmMicrokernelTester() |
| 7276 | .mr(3) |
| 7277 | .nr(8) |
| 7278 | .kr(4) |
| 7279 | .sr(1) |
| 7280 | .m(m) |
| 7281 | .n(n) |
| 7282 | .k(k) |
| 7283 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7284 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7285 | } |
| 7286 | } |
| 7287 | } |
| 7288 | } |
| 7289 | |
| 7290 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8) { |
| 7291 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7292 | for (size_t k = 16; k <= 80; k += 8) { |
| 7293 | GemmMicrokernelTester() |
| 7294 | .mr(3) |
| 7295 | .nr(8) |
| 7296 | .kr(4) |
| 7297 | .sr(1) |
| 7298 | .m(3) |
| 7299 | .n(8) |
| 7300 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7301 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7302 | } |
| 7303 | } |
| 7304 | |
| 7305 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8_strided_a) { |
| 7306 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7307 | for (size_t k = 16; k <= 80; k += 8) { |
| 7308 | GemmMicrokernelTester() |
| 7309 | .mr(3) |
| 7310 | .nr(8) |
| 7311 | .kr(4) |
| 7312 | .sr(1) |
| 7313 | .m(3) |
| 7314 | .n(8) |
| 7315 | .k(k) |
| 7316 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7317 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7318 | } |
| 7319 | } |
| 7320 | |
| 7321 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, k_div_8_subtile) { |
| 7322 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7323 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7324 | for (uint32_t n = 1; n <= 8; n++) { |
| 7325 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7326 | GemmMicrokernelTester() |
| 7327 | .mr(3) |
| 7328 | .nr(8) |
| 7329 | .kr(4) |
| 7330 | .sr(1) |
| 7331 | .m(m) |
| 7332 | .n(n) |
| 7333 | .k(k) |
| 7334 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7335 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7336 | } |
| 7337 | } |
| 7338 | } |
| 7339 | } |
| 7340 | |
| 7341 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8) { |
| 7342 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7343 | for (uint32_t n = 9; n < 16; n++) { |
| 7344 | for (size_t k = 1; k <= 40; k += 9) { |
| 7345 | GemmMicrokernelTester() |
| 7346 | .mr(3) |
| 7347 | .nr(8) |
| 7348 | .kr(4) |
| 7349 | .sr(1) |
| 7350 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7351 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7352 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7353 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7354 | } |
| 7355 | } |
| 7356 | } |
| 7357 | |
| 7358 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_strided_cn) { |
| 7359 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7360 | for (uint32_t n = 9; n < 16; n++) { |
| 7361 | for (size_t k = 1; k <= 40; k += 9) { |
| 7362 | GemmMicrokernelTester() |
| 7363 | .mr(3) |
| 7364 | .nr(8) |
| 7365 | .kr(4) |
| 7366 | .sr(1) |
| 7367 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7368 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7369 | .k(k) |
| 7370 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7371 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7372 | } |
| 7373 | } |
| 7374 | } |
| 7375 | |
| 7376 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_strided_a) { |
| 7377 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7378 | for (uint32_t n = 9; n < 16; n++) { |
| 7379 | for (size_t k = 1; k <= 40; k += 9) { |
| 7380 | GemmMicrokernelTester() |
| 7381 | .mr(3) |
| 7382 | .nr(8) |
| 7383 | .kr(4) |
| 7384 | .sr(1) |
| 7385 | .m(3) |
| 7386 | .n(n) |
| 7387 | .k(k) |
| 7388 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7389 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7390 | } |
| 7391 | } |
| 7392 | } |
| 7393 | |
| 7394 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_gt_8_subtile) { |
| 7395 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7396 | for (uint32_t n = 9; n < 16; n++) { |
| 7397 | for (size_t k = 1; k <= 40; k += 9) { |
| 7398 | for (uint32_t m = 1; m <= 3; m++) { |
| 7399 | GemmMicrokernelTester() |
| 7400 | .mr(3) |
| 7401 | .nr(8) |
| 7402 | .kr(4) |
| 7403 | .sr(1) |
| 7404 | .m(m) |
| 7405 | .n(n) |
| 7406 | .k(k) |
| 7407 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7408 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7409 | } |
| 7410 | } |
| 7411 | } |
| 7412 | } |
| 7413 | |
| 7414 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8) { |
| 7415 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7416 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7417 | for (size_t k = 1; k <= 40; k += 9) { |
| 7418 | GemmMicrokernelTester() |
| 7419 | .mr(3) |
| 7420 | .nr(8) |
| 7421 | .kr(4) |
| 7422 | .sr(1) |
| 7423 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7424 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7425 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7426 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7427 | } |
| 7428 | } |
| 7429 | } |
| 7430 | |
| 7431 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_strided_cn) { |
| 7432 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7433 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7434 | for (size_t k = 1; k <= 40; k += 9) { |
| 7435 | GemmMicrokernelTester() |
| 7436 | .mr(3) |
| 7437 | .nr(8) |
| 7438 | .kr(4) |
| 7439 | .sr(1) |
| 7440 | .m(3) |
| 7441 | .n(n) |
| 7442 | .k(k) |
| 7443 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7444 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7445 | } |
| 7446 | } |
| 7447 | } |
| 7448 | |
| 7449 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_strided_a) { |
| 7450 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7451 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7452 | for (size_t k = 1; k <= 40; k += 9) { |
| 7453 | GemmMicrokernelTester() |
| 7454 | .mr(3) |
| 7455 | .nr(8) |
| 7456 | .kr(4) |
| 7457 | .sr(1) |
| 7458 | .m(3) |
| 7459 | .n(n) |
| 7460 | .k(k) |
| 7461 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7462 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7463 | } |
| 7464 | } |
| 7465 | } |
| 7466 | |
| 7467 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, n_div_8_subtile) { |
| 7468 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7469 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7470 | for (size_t k = 1; k <= 40; k += 9) { |
| 7471 | for (uint32_t m = 1; m <= 3; m++) { |
| 7472 | GemmMicrokernelTester() |
| 7473 | .mr(3) |
| 7474 | .nr(8) |
| 7475 | .kr(4) |
| 7476 | .sr(1) |
| 7477 | .m(m) |
| 7478 | .n(n) |
| 7479 | .k(k) |
| 7480 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7481 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7482 | } |
| 7483 | } |
| 7484 | } |
| 7485 | } |
| 7486 | |
| 7487 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cm_subtile) { |
| 7488 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7489 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7490 | for (uint32_t n = 1; n <= 8; n++) { |
| 7491 | for (uint32_t m = 1; m <= 3; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7492 | GemmMicrokernelTester() |
| 7493 | .mr(3) |
| 7494 | .nr(8) |
| 7495 | .kr(4) |
| 7496 | .sr(1) |
| 7497 | .m(m) |
| 7498 | .n(n) |
| 7499 | .k(k) |
| 7500 | .cm_stride(11) |
| 7501 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7502 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7503 | } |
| 7504 | } |
| 7505 | } |
| 7506 | } |
| 7507 | |
| 7508 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, qmin) { |
| 7509 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7510 | GemmMicrokernelTester() |
| 7511 | .mr(3) |
| 7512 | .nr(8) |
| 7513 | .kr(4) |
| 7514 | .sr(1) |
| 7515 | .m(3) |
| 7516 | .n(8) |
| 7517 | .k(8) |
| 7518 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7519 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7520 | } |
| 7521 | |
| 7522 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, qmax) { |
| 7523 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7524 | GemmMicrokernelTester() |
| 7525 | .mr(3) |
| 7526 | .nr(8) |
| 7527 | .kr(4) |
| 7528 | .sr(1) |
| 7529 | .m(3) |
| 7530 | .n(8) |
| 7531 | .k(8) |
| 7532 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7533 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7534 | } |
| 7535 | |
| 7536 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, strided_cm) { |
| 7537 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7538 | GemmMicrokernelTester() |
| 7539 | .mr(3) |
| 7540 | .nr(8) |
| 7541 | .kr(4) |
| 7542 | .sr(1) |
| 7543 | .m(3) |
| 7544 | .n(8) |
| 7545 | .k(8) |
| 7546 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7547 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7548 | } |
| 7549 | |
| 7550 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_a_zero_point) { |
| 7551 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7552 | for (size_t k = 1; k <= 40; k += 9) { |
| 7553 | GemmMicrokernelTester() |
| 7554 | .mr(3) |
| 7555 | .nr(8) |
| 7556 | .kr(4) |
| 7557 | .sr(1) |
| 7558 | .m(3) |
| 7559 | .n(8) |
| 7560 | .k(k) |
| 7561 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7562 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7563 | } |
| 7564 | } |
| 7565 | |
| 7566 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_b_zero_point) { |
| 7567 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7568 | for (size_t k = 1; k <= 40; k += 9) { |
| 7569 | GemmMicrokernelTester() |
| 7570 | .mr(3) |
| 7571 | .nr(8) |
| 7572 | .kr(4) |
| 7573 | .sr(1) |
| 7574 | .m(3) |
| 7575 | .n(8) |
| 7576 | .k(k) |
| 7577 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7578 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7579 | } |
| 7580 | } |
| 7581 | |
| 7582 | TEST(QU8_GEMM_MINMAX_RNDNU_3X8C4__NEONDOT, no_zero_point) { |
| 7583 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7584 | for (size_t k = 1; k <= 40; k += 9) { |
| 7585 | GemmMicrokernelTester() |
| 7586 | .mr(3) |
| 7587 | .nr(8) |
| 7588 | .kr(4) |
| 7589 | .sr(1) |
| 7590 | .m(3) |
| 7591 | .n(8) |
| 7592 | .k(k) |
| 7593 | .a_zero_point(0) |
| 7594 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7595 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_3x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 7596 | } |
| 7597 | } |
| 7598 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 7599 | |
| 7600 | |
| 7601 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7602 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_eq_8) { |
| 7603 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7604 | GemmMicrokernelTester() |
| 7605 | .mr(4) |
| 7606 | .nr(8) |
| 7607 | .kr(4) |
| 7608 | .sr(1) |
| 7609 | .m(4) |
| 7610 | .n(8) |
| 7611 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7612 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7613 | } |
| 7614 | |
| 7615 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, strided_cn) { |
| 7616 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7617 | GemmMicrokernelTester() |
| 7618 | .mr(4) |
| 7619 | .nr(8) |
| 7620 | .kr(4) |
| 7621 | .sr(1) |
| 7622 | .m(4) |
| 7623 | .n(8) |
| 7624 | .k(8) |
| 7625 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7626 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7627 | } |
| 7628 | |
| 7629 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_eq_8_strided_a) { |
| 7630 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7631 | GemmMicrokernelTester() |
| 7632 | .mr(4) |
| 7633 | .nr(8) |
| 7634 | .kr(4) |
| 7635 | .sr(1) |
| 7636 | .m(4) |
| 7637 | .n(8) |
| 7638 | .k(8) |
| 7639 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7640 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7641 | } |
| 7642 | |
| 7643 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_eq_8_subtile) { |
| 7644 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7645 | for (uint32_t n = 1; n <= 8; n++) { |
| 7646 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7647 | GemmMicrokernelTester() |
| 7648 | .mr(4) |
| 7649 | .nr(8) |
| 7650 | .kr(4) |
| 7651 | .sr(1) |
| 7652 | .m(m) |
| 7653 | .n(n) |
| 7654 | .k(8) |
| 7655 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7656 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7657 | } |
| 7658 | } |
| 7659 | } |
| 7660 | |
| 7661 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_eq_8_subtile_m) { |
| 7662 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7663 | for (uint32_t m = 1; m <= 4; m++) { |
| 7664 | GemmMicrokernelTester() |
| 7665 | .mr(4) |
| 7666 | .nr(8) |
| 7667 | .kr(4) |
| 7668 | .sr(1) |
| 7669 | .m(m) |
| 7670 | .n(8) |
| 7671 | .k(8) |
| 7672 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7673 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7674 | } |
| 7675 | } |
| 7676 | |
| 7677 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_eq_8_subtile_n) { |
| 7678 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7679 | for (uint32_t n = 1; n <= 8; n++) { |
| 7680 | GemmMicrokernelTester() |
| 7681 | .mr(4) |
| 7682 | .nr(8) |
| 7683 | .kr(4) |
| 7684 | .sr(1) |
| 7685 | .m(4) |
| 7686 | .n(n) |
| 7687 | .k(8) |
| 7688 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7689 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7690 | } |
| 7691 | } |
| 7692 | |
| 7693 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_lt_8) { |
| 7694 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7695 | for (size_t k = 1; k < 8; k++) { |
| 7696 | GemmMicrokernelTester() |
| 7697 | .mr(4) |
| 7698 | .nr(8) |
| 7699 | .kr(4) |
| 7700 | .sr(1) |
| 7701 | .m(4) |
| 7702 | .n(8) |
| 7703 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7704 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7705 | } |
| 7706 | } |
| 7707 | |
| 7708 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_lt_8_strided_a) { |
| 7709 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7710 | for (size_t k = 1; k < 8; k++) { |
| 7711 | GemmMicrokernelTester() |
| 7712 | .mr(4) |
| 7713 | .nr(8) |
| 7714 | .kr(4) |
| 7715 | .sr(1) |
| 7716 | .m(4) |
| 7717 | .n(8) |
| 7718 | .k(k) |
| 7719 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7720 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7721 | } |
| 7722 | } |
| 7723 | |
| 7724 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_lt_8_subtile) { |
| 7725 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7726 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7727 | for (uint32_t n = 1; n <= 8; n++) { |
| 7728 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7729 | GemmMicrokernelTester() |
| 7730 | .mr(4) |
| 7731 | .nr(8) |
| 7732 | .kr(4) |
| 7733 | .sr(1) |
| 7734 | .m(m) |
| 7735 | .n(n) |
| 7736 | .k(k) |
| 7737 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7738 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7739 | } |
| 7740 | } |
| 7741 | } |
| 7742 | } |
| 7743 | |
| 7744 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_gt_8) { |
| 7745 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7746 | for (size_t k = 9; k < 16; k++) { |
| 7747 | GemmMicrokernelTester() |
| 7748 | .mr(4) |
| 7749 | .nr(8) |
| 7750 | .kr(4) |
| 7751 | .sr(1) |
| 7752 | .m(4) |
| 7753 | .n(8) |
| 7754 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7755 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7756 | } |
| 7757 | } |
| 7758 | |
| 7759 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_gt_8_strided_a) { |
| 7760 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7761 | for (size_t k = 9; k < 16; k++) { |
| 7762 | GemmMicrokernelTester() |
| 7763 | .mr(4) |
| 7764 | .nr(8) |
| 7765 | .kr(4) |
| 7766 | .sr(1) |
| 7767 | .m(4) |
| 7768 | .n(8) |
| 7769 | .k(k) |
| 7770 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7771 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7772 | } |
| 7773 | } |
| 7774 | |
| 7775 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_gt_8_subtile) { |
| 7776 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7777 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7778 | for (uint32_t n = 1; n <= 8; n++) { |
| 7779 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7780 | GemmMicrokernelTester() |
| 7781 | .mr(4) |
| 7782 | .nr(8) |
| 7783 | .kr(4) |
| 7784 | .sr(1) |
| 7785 | .m(m) |
| 7786 | .n(n) |
| 7787 | .k(k) |
| 7788 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7789 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7790 | } |
| 7791 | } |
| 7792 | } |
| 7793 | } |
| 7794 | |
| 7795 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_div_8) { |
| 7796 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7797 | for (size_t k = 16; k <= 80; k += 8) { |
| 7798 | GemmMicrokernelTester() |
| 7799 | .mr(4) |
| 7800 | .nr(8) |
| 7801 | .kr(4) |
| 7802 | .sr(1) |
| 7803 | .m(4) |
| 7804 | .n(8) |
| 7805 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7806 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7807 | } |
| 7808 | } |
| 7809 | |
| 7810 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_div_8_strided_a) { |
| 7811 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7812 | for (size_t k = 16; k <= 80; k += 8) { |
| 7813 | GemmMicrokernelTester() |
| 7814 | .mr(4) |
| 7815 | .nr(8) |
| 7816 | .kr(4) |
| 7817 | .sr(1) |
| 7818 | .m(4) |
| 7819 | .n(8) |
| 7820 | .k(k) |
| 7821 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7822 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7823 | } |
| 7824 | } |
| 7825 | |
| 7826 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, k_div_8_subtile) { |
| 7827 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7828 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7829 | for (uint32_t n = 1; n <= 8; n++) { |
| 7830 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7831 | GemmMicrokernelTester() |
| 7832 | .mr(4) |
| 7833 | .nr(8) |
| 7834 | .kr(4) |
| 7835 | .sr(1) |
| 7836 | .m(m) |
| 7837 | .n(n) |
| 7838 | .k(k) |
| 7839 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7840 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7841 | } |
| 7842 | } |
| 7843 | } |
| 7844 | } |
| 7845 | |
| 7846 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_gt_8) { |
| 7847 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7848 | for (uint32_t n = 9; n < 16; n++) { |
| 7849 | for (size_t k = 1; k <= 40; k += 9) { |
| 7850 | GemmMicrokernelTester() |
| 7851 | .mr(4) |
| 7852 | .nr(8) |
| 7853 | .kr(4) |
| 7854 | .sr(1) |
| 7855 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7856 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7857 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7858 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7859 | } |
| 7860 | } |
| 7861 | } |
| 7862 | |
| 7863 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_gt_8_strided_cn) { |
| 7864 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7865 | for (uint32_t n = 9; n < 16; n++) { |
| 7866 | for (size_t k = 1; k <= 40; k += 9) { |
| 7867 | GemmMicrokernelTester() |
| 7868 | .mr(4) |
| 7869 | .nr(8) |
| 7870 | .kr(4) |
| 7871 | .sr(1) |
| 7872 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7873 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7874 | .k(k) |
| 7875 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7876 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7877 | } |
| 7878 | } |
| 7879 | } |
| 7880 | |
| 7881 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_gt_8_strided_a) { |
| 7882 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7883 | for (uint32_t n = 9; n < 16; n++) { |
| 7884 | for (size_t k = 1; k <= 40; k += 9) { |
| 7885 | GemmMicrokernelTester() |
| 7886 | .mr(4) |
| 7887 | .nr(8) |
| 7888 | .kr(4) |
| 7889 | .sr(1) |
| 7890 | .m(4) |
| 7891 | .n(n) |
| 7892 | .k(k) |
| 7893 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7894 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7895 | } |
| 7896 | } |
| 7897 | } |
| 7898 | |
| 7899 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_gt_8_subtile) { |
| 7900 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7901 | for (uint32_t n = 9; n < 16; n++) { |
| 7902 | for (size_t k = 1; k <= 40; k += 9) { |
| 7903 | for (uint32_t m = 1; m <= 4; m++) { |
| 7904 | GemmMicrokernelTester() |
| 7905 | .mr(4) |
| 7906 | .nr(8) |
| 7907 | .kr(4) |
| 7908 | .sr(1) |
| 7909 | .m(m) |
| 7910 | .n(n) |
| 7911 | .k(k) |
| 7912 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7913 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7914 | } |
| 7915 | } |
| 7916 | } |
| 7917 | } |
| 7918 | |
| 7919 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_div_8) { |
| 7920 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7921 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7922 | for (size_t k = 1; k <= 40; k += 9) { |
| 7923 | GemmMicrokernelTester() |
| 7924 | .mr(4) |
| 7925 | .nr(8) |
| 7926 | .kr(4) |
| 7927 | .sr(1) |
| 7928 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7929 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7930 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7931 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7932 | } |
| 7933 | } |
| 7934 | } |
| 7935 | |
| 7936 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_div_8_strided_cn) { |
| 7937 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7938 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7939 | for (size_t k = 1; k <= 40; k += 9) { |
| 7940 | GemmMicrokernelTester() |
| 7941 | .mr(4) |
| 7942 | .nr(8) |
| 7943 | .kr(4) |
| 7944 | .sr(1) |
| 7945 | .m(4) |
| 7946 | .n(n) |
| 7947 | .k(k) |
| 7948 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7949 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7950 | } |
| 7951 | } |
| 7952 | } |
| 7953 | |
| 7954 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_div_8_strided_a) { |
| 7955 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7956 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7957 | for (size_t k = 1; k <= 40; k += 9) { |
| 7958 | GemmMicrokernelTester() |
| 7959 | .mr(4) |
| 7960 | .nr(8) |
| 7961 | .kr(4) |
| 7962 | .sr(1) |
| 7963 | .m(4) |
| 7964 | .n(n) |
| 7965 | .k(k) |
| 7966 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7967 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7968 | } |
| 7969 | } |
| 7970 | } |
| 7971 | |
| 7972 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, n_div_8_subtile) { |
| 7973 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7974 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7975 | for (size_t k = 1; k <= 40; k += 9) { |
| 7976 | for (uint32_t m = 1; m <= 4; m++) { |
| 7977 | GemmMicrokernelTester() |
| 7978 | .mr(4) |
| 7979 | .nr(8) |
| 7980 | .kr(4) |
| 7981 | .sr(1) |
| 7982 | .m(m) |
| 7983 | .n(n) |
| 7984 | .k(k) |
| 7985 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7986 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7987 | } |
| 7988 | } |
| 7989 | } |
| 7990 | } |
| 7991 | |
| 7992 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, strided_cm_subtile) { |
| 7993 | TEST_REQUIRES_ARM_NEON_DOT; |
| 7994 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7995 | for (uint32_t n = 1; n <= 8; n++) { |
| 7996 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 7997 | GemmMicrokernelTester() |
| 7998 | .mr(4) |
| 7999 | .nr(8) |
| 8000 | .kr(4) |
| 8001 | .sr(1) |
| 8002 | .m(m) |
| 8003 | .n(n) |
| 8004 | .k(k) |
| 8005 | .cm_stride(11) |
| 8006 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8007 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8008 | } |
| 8009 | } |
| 8010 | } |
| 8011 | } |
| 8012 | |
| 8013 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, qmin) { |
| 8014 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8015 | GemmMicrokernelTester() |
| 8016 | .mr(4) |
| 8017 | .nr(8) |
| 8018 | .kr(4) |
| 8019 | .sr(1) |
| 8020 | .m(4) |
| 8021 | .n(8) |
| 8022 | .k(8) |
| 8023 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8024 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8025 | } |
| 8026 | |
| 8027 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, qmax) { |
| 8028 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8029 | GemmMicrokernelTester() |
| 8030 | .mr(4) |
| 8031 | .nr(8) |
| 8032 | .kr(4) |
| 8033 | .sr(1) |
| 8034 | .m(4) |
| 8035 | .n(8) |
| 8036 | .k(8) |
| 8037 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8038 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8039 | } |
| 8040 | |
| 8041 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, strided_cm) { |
| 8042 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8043 | GemmMicrokernelTester() |
| 8044 | .mr(4) |
| 8045 | .nr(8) |
| 8046 | .kr(4) |
| 8047 | .sr(1) |
| 8048 | .m(4) |
| 8049 | .n(8) |
| 8050 | .k(8) |
| 8051 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8052 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8053 | } |
| 8054 | |
| 8055 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, no_a_zero_point) { |
| 8056 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8057 | for (size_t k = 1; k <= 40; k += 9) { |
| 8058 | GemmMicrokernelTester() |
| 8059 | .mr(4) |
| 8060 | .nr(8) |
| 8061 | .kr(4) |
| 8062 | .sr(1) |
| 8063 | .m(4) |
| 8064 | .n(8) |
| 8065 | .k(k) |
| 8066 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8067 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8068 | } |
| 8069 | } |
| 8070 | |
| 8071 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, no_b_zero_point) { |
| 8072 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8073 | for (size_t k = 1; k <= 40; k += 9) { |
| 8074 | GemmMicrokernelTester() |
| 8075 | .mr(4) |
| 8076 | .nr(8) |
| 8077 | .kr(4) |
| 8078 | .sr(1) |
| 8079 | .m(4) |
| 8080 | .n(8) |
| 8081 | .k(k) |
| 8082 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8083 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8084 | } |
| 8085 | } |
| 8086 | |
| 8087 | TEST(QU8_GEMM_MINMAX_RNDNU_4X8C4__NEONDOT, no_zero_point) { |
| 8088 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8089 | for (size_t k = 1; k <= 40; k += 9) { |
| 8090 | GemmMicrokernelTester() |
| 8091 | .mr(4) |
| 8092 | .nr(8) |
| 8093 | .kr(4) |
| 8094 | .sr(1) |
| 8095 | .m(4) |
| 8096 | .n(8) |
| 8097 | .k(k) |
| 8098 | .a_zero_point(0) |
| 8099 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8100 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8101 | } |
| 8102 | } |
| 8103 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 8104 | |
| 8105 | |
| 8106 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8107 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8) { |
| 8108 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8109 | GemmMicrokernelTester() |
| 8110 | .mr(5) |
| 8111 | .nr(8) |
| 8112 | .kr(4) |
| 8113 | .sr(1) |
| 8114 | .m(5) |
| 8115 | .n(8) |
| 8116 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8117 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8118 | } |
| 8119 | |
| 8120 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cn) { |
| 8121 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8122 | GemmMicrokernelTester() |
| 8123 | .mr(5) |
| 8124 | .nr(8) |
| 8125 | .kr(4) |
| 8126 | .sr(1) |
| 8127 | .m(5) |
| 8128 | .n(8) |
| 8129 | .k(8) |
| 8130 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8131 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8132 | } |
| 8133 | |
| 8134 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_strided_a) { |
| 8135 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8136 | GemmMicrokernelTester() |
| 8137 | .mr(5) |
| 8138 | .nr(8) |
| 8139 | .kr(4) |
| 8140 | .sr(1) |
| 8141 | .m(5) |
| 8142 | .n(8) |
| 8143 | .k(8) |
| 8144 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8145 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8146 | } |
| 8147 | |
| 8148 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile) { |
| 8149 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8150 | for (uint32_t n = 1; n <= 8; n++) { |
| 8151 | for (uint32_t m = 1; m <= 5; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8152 | GemmMicrokernelTester() |
| 8153 | .mr(5) |
| 8154 | .nr(8) |
| 8155 | .kr(4) |
| 8156 | .sr(1) |
| 8157 | .m(m) |
| 8158 | .n(n) |
| 8159 | .k(8) |
| 8160 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8161 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8162 | } |
| 8163 | } |
| 8164 | } |
| 8165 | |
| 8166 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile_m) { |
| 8167 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8168 | for (uint32_t m = 1; m <= 5; m++) { |
| 8169 | GemmMicrokernelTester() |
| 8170 | .mr(5) |
| 8171 | .nr(8) |
| 8172 | .kr(4) |
| 8173 | .sr(1) |
| 8174 | .m(m) |
| 8175 | .n(8) |
| 8176 | .k(8) |
| 8177 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8178 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8179 | } |
| 8180 | } |
| 8181 | |
| 8182 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_eq_8_subtile_n) { |
| 8183 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8184 | for (uint32_t n = 1; n <= 8; n++) { |
| 8185 | GemmMicrokernelTester() |
| 8186 | .mr(5) |
| 8187 | .nr(8) |
| 8188 | .kr(4) |
| 8189 | .sr(1) |
| 8190 | .m(5) |
| 8191 | .n(n) |
| 8192 | .k(8) |
| 8193 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8194 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8195 | } |
| 8196 | } |
| 8197 | |
| 8198 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8) { |
| 8199 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8200 | for (size_t k = 1; k < 8; k++) { |
| 8201 | GemmMicrokernelTester() |
| 8202 | .mr(5) |
| 8203 | .nr(8) |
| 8204 | .kr(4) |
| 8205 | .sr(1) |
| 8206 | .m(5) |
| 8207 | .n(8) |
| 8208 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8209 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8210 | } |
| 8211 | } |
| 8212 | |
| 8213 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8_strided_a) { |
| 8214 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8215 | for (size_t k = 1; k < 8; k++) { |
| 8216 | GemmMicrokernelTester() |
| 8217 | .mr(5) |
| 8218 | .nr(8) |
| 8219 | .kr(4) |
| 8220 | .sr(1) |
| 8221 | .m(5) |
| 8222 | .n(8) |
| 8223 | .k(k) |
| 8224 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8225 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8226 | } |
| 8227 | } |
| 8228 | |
| 8229 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_lt_8_subtile) { |
| 8230 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8231 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8232 | for (uint32_t n = 1; n <= 8; n++) { |
| 8233 | for (uint32_t m = 1; m <= 5; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8234 | GemmMicrokernelTester() |
| 8235 | .mr(5) |
| 8236 | .nr(8) |
| 8237 | .kr(4) |
| 8238 | .sr(1) |
| 8239 | .m(m) |
| 8240 | .n(n) |
| 8241 | .k(k) |
| 8242 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8243 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8244 | } |
| 8245 | } |
| 8246 | } |
| 8247 | } |
| 8248 | |
| 8249 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8) { |
| 8250 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8251 | for (size_t k = 9; k < 16; k++) { |
| 8252 | GemmMicrokernelTester() |
| 8253 | .mr(5) |
| 8254 | .nr(8) |
| 8255 | .kr(4) |
| 8256 | .sr(1) |
| 8257 | .m(5) |
| 8258 | .n(8) |
| 8259 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8260 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8261 | } |
| 8262 | } |
| 8263 | |
| 8264 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8_strided_a) { |
| 8265 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8266 | for (size_t k = 9; k < 16; k++) { |
| 8267 | GemmMicrokernelTester() |
| 8268 | .mr(5) |
| 8269 | .nr(8) |
| 8270 | .kr(4) |
| 8271 | .sr(1) |
| 8272 | .m(5) |
| 8273 | .n(8) |
| 8274 | .k(k) |
| 8275 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8276 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8277 | } |
| 8278 | } |
| 8279 | |
| 8280 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_gt_8_subtile) { |
| 8281 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8282 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8283 | for (uint32_t n = 1; n <= 8; n++) { |
| 8284 | for (uint32_t m = 1; m <= 5; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8285 | GemmMicrokernelTester() |
| 8286 | .mr(5) |
| 8287 | .nr(8) |
| 8288 | .kr(4) |
| 8289 | .sr(1) |
| 8290 | .m(m) |
| 8291 | .n(n) |
| 8292 | .k(k) |
| 8293 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8294 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8295 | } |
| 8296 | } |
| 8297 | } |
| 8298 | } |
| 8299 | |
| 8300 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8) { |
| 8301 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8302 | for (size_t k = 16; k <= 80; k += 8) { |
| 8303 | GemmMicrokernelTester() |
| 8304 | .mr(5) |
| 8305 | .nr(8) |
| 8306 | .kr(4) |
| 8307 | .sr(1) |
| 8308 | .m(5) |
| 8309 | .n(8) |
| 8310 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8311 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8312 | } |
| 8313 | } |
| 8314 | |
| 8315 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8_strided_a) { |
| 8316 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8317 | for (size_t k = 16; k <= 80; k += 8) { |
| 8318 | GemmMicrokernelTester() |
| 8319 | .mr(5) |
| 8320 | .nr(8) |
| 8321 | .kr(4) |
| 8322 | .sr(1) |
| 8323 | .m(5) |
| 8324 | .n(8) |
| 8325 | .k(k) |
| 8326 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8327 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8328 | } |
| 8329 | } |
| 8330 | |
| 8331 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, k_div_8_subtile) { |
| 8332 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8333 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8334 | for (uint32_t n = 1; n <= 8; n++) { |
| 8335 | for (uint32_t m = 1; m <= 5; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8336 | GemmMicrokernelTester() |
| 8337 | .mr(5) |
| 8338 | .nr(8) |
| 8339 | .kr(4) |
| 8340 | .sr(1) |
| 8341 | .m(m) |
| 8342 | .n(n) |
| 8343 | .k(k) |
| 8344 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8345 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8346 | } |
| 8347 | } |
| 8348 | } |
| 8349 | } |
| 8350 | |
| 8351 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8) { |
| 8352 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8353 | for (uint32_t n = 9; n < 16; n++) { |
| 8354 | for (size_t k = 1; k <= 40; k += 9) { |
| 8355 | GemmMicrokernelTester() |
| 8356 | .mr(5) |
| 8357 | .nr(8) |
| 8358 | .kr(4) |
| 8359 | .sr(1) |
| 8360 | .m(5) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8361 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8362 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8363 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8364 | } |
| 8365 | } |
| 8366 | } |
| 8367 | |
| 8368 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_strided_cn) { |
| 8369 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8370 | for (uint32_t n = 9; n < 16; n++) { |
| 8371 | for (size_t k = 1; k <= 40; k += 9) { |
| 8372 | GemmMicrokernelTester() |
| 8373 | .mr(5) |
| 8374 | .nr(8) |
| 8375 | .kr(4) |
| 8376 | .sr(1) |
| 8377 | .m(5) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8378 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8379 | .k(k) |
| 8380 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8381 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8382 | } |
| 8383 | } |
| 8384 | } |
| 8385 | |
| 8386 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_strided_a) { |
| 8387 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8388 | for (uint32_t n = 9; n < 16; n++) { |
| 8389 | for (size_t k = 1; k <= 40; k += 9) { |
| 8390 | GemmMicrokernelTester() |
| 8391 | .mr(5) |
| 8392 | .nr(8) |
| 8393 | .kr(4) |
| 8394 | .sr(1) |
| 8395 | .m(5) |
| 8396 | .n(n) |
| 8397 | .k(k) |
| 8398 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8399 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8400 | } |
| 8401 | } |
| 8402 | } |
| 8403 | |
| 8404 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_gt_8_subtile) { |
| 8405 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8406 | for (uint32_t n = 9; n < 16; n++) { |
| 8407 | for (size_t k = 1; k <= 40; k += 9) { |
| 8408 | for (uint32_t m = 1; m <= 5; m++) { |
| 8409 | GemmMicrokernelTester() |
| 8410 | .mr(5) |
| 8411 | .nr(8) |
| 8412 | .kr(4) |
| 8413 | .sr(1) |
| 8414 | .m(m) |
| 8415 | .n(n) |
| 8416 | .k(k) |
| 8417 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8418 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8419 | } |
| 8420 | } |
| 8421 | } |
| 8422 | } |
| 8423 | |
| 8424 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8) { |
| 8425 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8426 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8427 | for (size_t k = 1; k <= 40; k += 9) { |
| 8428 | GemmMicrokernelTester() |
| 8429 | .mr(5) |
| 8430 | .nr(8) |
| 8431 | .kr(4) |
| 8432 | .sr(1) |
| 8433 | .m(5) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8434 | .n(n) |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8435 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8436 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8437 | } |
| 8438 | } |
| 8439 | } |
| 8440 | |
| 8441 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_strided_cn) { |
| 8442 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8443 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8444 | for (size_t k = 1; k <= 40; k += 9) { |
| 8445 | GemmMicrokernelTester() |
| 8446 | .mr(5) |
| 8447 | .nr(8) |
| 8448 | .kr(4) |
| 8449 | .sr(1) |
| 8450 | .m(5) |
| 8451 | .n(n) |
| 8452 | .k(k) |
| 8453 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8454 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8455 | } |
| 8456 | } |
| 8457 | } |
| 8458 | |
| 8459 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_strided_a) { |
| 8460 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8461 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8462 | for (size_t k = 1; k <= 40; k += 9) { |
| 8463 | GemmMicrokernelTester() |
| 8464 | .mr(5) |
| 8465 | .nr(8) |
| 8466 | .kr(4) |
| 8467 | .sr(1) |
| 8468 | .m(5) |
| 8469 | .n(n) |
| 8470 | .k(k) |
| 8471 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8472 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8473 | } |
| 8474 | } |
| 8475 | } |
| 8476 | |
| 8477 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, n_div_8_subtile) { |
| 8478 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8479 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8480 | for (size_t k = 1; k <= 40; k += 9) { |
| 8481 | for (uint32_t m = 1; m <= 5; m++) { |
| 8482 | GemmMicrokernelTester() |
| 8483 | .mr(5) |
| 8484 | .nr(8) |
| 8485 | .kr(4) |
| 8486 | .sr(1) |
| 8487 | .m(m) |
| 8488 | .n(n) |
| 8489 | .k(k) |
| 8490 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8491 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8492 | } |
| 8493 | } |
| 8494 | } |
| 8495 | } |
| 8496 | |
| 8497 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cm_subtile) { |
| 8498 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8499 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8500 | for (uint32_t n = 1; n <= 8; n++) { |
| 8501 | for (uint32_t m = 1; m <= 5; m++) { |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8502 | GemmMicrokernelTester() |
| 8503 | .mr(5) |
| 8504 | .nr(8) |
| 8505 | .kr(4) |
| 8506 | .sr(1) |
| 8507 | .m(m) |
| 8508 | .n(n) |
| 8509 | .k(k) |
| 8510 | .cm_stride(11) |
| 8511 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8512 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8513 | } |
| 8514 | } |
| 8515 | } |
| 8516 | } |
| 8517 | |
| 8518 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, qmin) { |
| 8519 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8520 | GemmMicrokernelTester() |
| 8521 | .mr(5) |
| 8522 | .nr(8) |
| 8523 | .kr(4) |
| 8524 | .sr(1) |
| 8525 | .m(5) |
| 8526 | .n(8) |
| 8527 | .k(8) |
| 8528 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8529 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8530 | } |
| 8531 | |
| 8532 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, qmax) { |
| 8533 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8534 | GemmMicrokernelTester() |
| 8535 | .mr(5) |
| 8536 | .nr(8) |
| 8537 | .kr(4) |
| 8538 | .sr(1) |
| 8539 | .m(5) |
| 8540 | .n(8) |
| 8541 | .k(8) |
| 8542 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8543 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8544 | } |
| 8545 | |
| 8546 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, strided_cm) { |
| 8547 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8548 | GemmMicrokernelTester() |
| 8549 | .mr(5) |
| 8550 | .nr(8) |
| 8551 | .kr(4) |
| 8552 | .sr(1) |
| 8553 | .m(5) |
| 8554 | .n(8) |
| 8555 | .k(8) |
| 8556 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8557 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8558 | } |
| 8559 | |
| 8560 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_a_zero_point) { |
| 8561 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8562 | for (size_t k = 1; k <= 40; k += 9) { |
| 8563 | GemmMicrokernelTester() |
| 8564 | .mr(5) |
| 8565 | .nr(8) |
| 8566 | .kr(4) |
| 8567 | .sr(1) |
| 8568 | .m(5) |
| 8569 | .n(8) |
| 8570 | .k(k) |
| 8571 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8572 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8573 | } |
| 8574 | } |
| 8575 | |
| 8576 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_b_zero_point) { |
| 8577 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8578 | for (size_t k = 1; k <= 40; k += 9) { |
| 8579 | GemmMicrokernelTester() |
| 8580 | .mr(5) |
| 8581 | .nr(8) |
| 8582 | .kr(4) |
| 8583 | .sr(1) |
| 8584 | .m(5) |
| 8585 | .n(8) |
| 8586 | .k(k) |
| 8587 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8588 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8589 | } |
| 8590 | } |
| 8591 | |
| 8592 | TEST(QU8_GEMM_MINMAX_RNDNU_5X8C4__NEONDOT, no_zero_point) { |
| 8593 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8594 | for (size_t k = 1; k <= 40; k += 9) { |
| 8595 | GemmMicrokernelTester() |
| 8596 | .mr(5) |
| 8597 | .nr(8) |
| 8598 | .kr(4) |
| 8599 | .sr(1) |
| 8600 | .m(5) |
| 8601 | .n(8) |
| 8602 | .k(k) |
| 8603 | .a_zero_point(0) |
| 8604 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8605 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_5x8c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | e033126 | 2021-08-11 23:18:59 -0700 | [diff] [blame] | 8606 | } |
| 8607 | } |
| 8608 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 8609 | |
| 8610 | |
| 8611 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8612 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_eq_8) { |
| 8613 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8614 | GemmMicrokernelTester() |
| 8615 | .mr(1) |
| 8616 | .nr(16) |
| 8617 | .kr(4) |
| 8618 | .sr(1) |
| 8619 | .m(1) |
| 8620 | .n(16) |
| 8621 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8622 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8623 | } |
| 8624 | |
| 8625 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, strided_cn) { |
| 8626 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8627 | GemmMicrokernelTester() |
| 8628 | .mr(1) |
| 8629 | .nr(16) |
| 8630 | .kr(4) |
| 8631 | .sr(1) |
| 8632 | .m(1) |
| 8633 | .n(16) |
| 8634 | .k(8) |
| 8635 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8636 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8637 | } |
| 8638 | |
| 8639 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_eq_8_strided_a) { |
| 8640 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8641 | GemmMicrokernelTester() |
| 8642 | .mr(1) |
| 8643 | .nr(16) |
| 8644 | .kr(4) |
| 8645 | .sr(1) |
| 8646 | .m(1) |
| 8647 | .n(16) |
| 8648 | .k(8) |
| 8649 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8650 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8651 | } |
| 8652 | |
| 8653 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_eq_8_subtile) { |
| 8654 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8655 | for (uint32_t n = 1; n <= 16; n++) { |
| 8656 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8657 | GemmMicrokernelTester() |
| 8658 | .mr(1) |
| 8659 | .nr(16) |
| 8660 | .kr(4) |
| 8661 | .sr(1) |
| 8662 | .m(m) |
| 8663 | .n(n) |
| 8664 | .k(8) |
| 8665 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8666 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8667 | } |
| 8668 | } |
| 8669 | } |
| 8670 | |
| 8671 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_eq_8_subtile_m) { |
| 8672 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8673 | for (uint32_t m = 1; m <= 1; m++) { |
| 8674 | GemmMicrokernelTester() |
| 8675 | .mr(1) |
| 8676 | .nr(16) |
| 8677 | .kr(4) |
| 8678 | .sr(1) |
| 8679 | .m(m) |
| 8680 | .n(16) |
| 8681 | .k(8) |
| 8682 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8683 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8684 | } |
| 8685 | } |
| 8686 | |
| 8687 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_eq_8_subtile_n) { |
| 8688 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8689 | for (uint32_t n = 1; n <= 16; n++) { |
| 8690 | GemmMicrokernelTester() |
| 8691 | .mr(1) |
| 8692 | .nr(16) |
| 8693 | .kr(4) |
| 8694 | .sr(1) |
| 8695 | .m(1) |
| 8696 | .n(n) |
| 8697 | .k(8) |
| 8698 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8699 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8700 | } |
| 8701 | } |
| 8702 | |
| 8703 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_lt_8) { |
| 8704 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8705 | for (size_t k = 1; k < 8; k++) { |
| 8706 | GemmMicrokernelTester() |
| 8707 | .mr(1) |
| 8708 | .nr(16) |
| 8709 | .kr(4) |
| 8710 | .sr(1) |
| 8711 | .m(1) |
| 8712 | .n(16) |
| 8713 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8714 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8715 | } |
| 8716 | } |
| 8717 | |
| 8718 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_lt_8_strided_a) { |
| 8719 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8720 | for (size_t k = 1; k < 8; k++) { |
| 8721 | GemmMicrokernelTester() |
| 8722 | .mr(1) |
| 8723 | .nr(16) |
| 8724 | .kr(4) |
| 8725 | .sr(1) |
| 8726 | .m(1) |
| 8727 | .n(16) |
| 8728 | .k(k) |
| 8729 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8730 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8731 | } |
| 8732 | } |
| 8733 | |
| 8734 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_lt_8_subtile) { |
| 8735 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8736 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8737 | for (uint32_t n = 1; n <= 16; n++) { |
| 8738 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8739 | GemmMicrokernelTester() |
| 8740 | .mr(1) |
| 8741 | .nr(16) |
| 8742 | .kr(4) |
| 8743 | .sr(1) |
| 8744 | .m(m) |
| 8745 | .n(n) |
| 8746 | .k(k) |
| 8747 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8748 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8749 | } |
| 8750 | } |
| 8751 | } |
| 8752 | } |
| 8753 | |
| 8754 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_gt_8) { |
| 8755 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8756 | for (size_t k = 9; k < 16; k++) { |
| 8757 | GemmMicrokernelTester() |
| 8758 | .mr(1) |
| 8759 | .nr(16) |
| 8760 | .kr(4) |
| 8761 | .sr(1) |
| 8762 | .m(1) |
| 8763 | .n(16) |
| 8764 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8765 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8766 | } |
| 8767 | } |
| 8768 | |
| 8769 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_gt_8_strided_a) { |
| 8770 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8771 | for (size_t k = 9; k < 16; k++) { |
| 8772 | GemmMicrokernelTester() |
| 8773 | .mr(1) |
| 8774 | .nr(16) |
| 8775 | .kr(4) |
| 8776 | .sr(1) |
| 8777 | .m(1) |
| 8778 | .n(16) |
| 8779 | .k(k) |
| 8780 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8781 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8782 | } |
| 8783 | } |
| 8784 | |
| 8785 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_gt_8_subtile) { |
| 8786 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8787 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8788 | for (uint32_t n = 1; n <= 16; n++) { |
| 8789 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8790 | GemmMicrokernelTester() |
| 8791 | .mr(1) |
| 8792 | .nr(16) |
| 8793 | .kr(4) |
| 8794 | .sr(1) |
| 8795 | .m(m) |
| 8796 | .n(n) |
| 8797 | .k(k) |
| 8798 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8799 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8800 | } |
| 8801 | } |
| 8802 | } |
| 8803 | } |
| 8804 | |
| 8805 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_div_8) { |
| 8806 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8807 | for (size_t k = 16; k <= 80; k += 8) { |
| 8808 | GemmMicrokernelTester() |
| 8809 | .mr(1) |
| 8810 | .nr(16) |
| 8811 | .kr(4) |
| 8812 | .sr(1) |
| 8813 | .m(1) |
| 8814 | .n(16) |
| 8815 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8816 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8817 | } |
| 8818 | } |
| 8819 | |
| 8820 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_div_8_strided_a) { |
| 8821 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8822 | for (size_t k = 16; k <= 80; k += 8) { |
| 8823 | GemmMicrokernelTester() |
| 8824 | .mr(1) |
| 8825 | .nr(16) |
| 8826 | .kr(4) |
| 8827 | .sr(1) |
| 8828 | .m(1) |
| 8829 | .n(16) |
| 8830 | .k(k) |
| 8831 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8832 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8833 | } |
| 8834 | } |
| 8835 | |
| 8836 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, k_div_8_subtile) { |
| 8837 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8838 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8839 | for (uint32_t n = 1; n <= 16; n++) { |
| 8840 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8841 | GemmMicrokernelTester() |
| 8842 | .mr(1) |
| 8843 | .nr(16) |
| 8844 | .kr(4) |
| 8845 | .sr(1) |
| 8846 | .m(m) |
| 8847 | .n(n) |
| 8848 | .k(k) |
| 8849 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8850 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8851 | } |
| 8852 | } |
| 8853 | } |
| 8854 | } |
| 8855 | |
| 8856 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_gt_16) { |
| 8857 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8858 | for (uint32_t n = 17; n < 32; n++) { |
| 8859 | for (size_t k = 1; k <= 40; k += 9) { |
| 8860 | GemmMicrokernelTester() |
| 8861 | .mr(1) |
| 8862 | .nr(16) |
| 8863 | .kr(4) |
| 8864 | .sr(1) |
| 8865 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8866 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8867 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8868 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8869 | } |
| 8870 | } |
| 8871 | } |
| 8872 | |
| 8873 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_gt_16_strided_cn) { |
| 8874 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8875 | for (uint32_t n = 17; n < 32; n++) { |
| 8876 | for (size_t k = 1; k <= 40; k += 9) { |
| 8877 | GemmMicrokernelTester() |
| 8878 | .mr(1) |
| 8879 | .nr(16) |
| 8880 | .kr(4) |
| 8881 | .sr(1) |
| 8882 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8883 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8884 | .k(k) |
| 8885 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8886 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8887 | } |
| 8888 | } |
| 8889 | } |
| 8890 | |
| 8891 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_gt_16_strided_a) { |
| 8892 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8893 | for (uint32_t n = 17; n < 32; n++) { |
| 8894 | for (size_t k = 1; k <= 40; k += 9) { |
| 8895 | GemmMicrokernelTester() |
| 8896 | .mr(1) |
| 8897 | .nr(16) |
| 8898 | .kr(4) |
| 8899 | .sr(1) |
| 8900 | .m(1) |
| 8901 | .n(n) |
| 8902 | .k(k) |
| 8903 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8904 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8905 | } |
| 8906 | } |
| 8907 | } |
| 8908 | |
| 8909 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_gt_16_subtile) { |
| 8910 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8911 | for (uint32_t n = 17; n < 32; n++) { |
| 8912 | for (size_t k = 1; k <= 40; k += 9) { |
| 8913 | for (uint32_t m = 1; m <= 1; m++) { |
| 8914 | GemmMicrokernelTester() |
| 8915 | .mr(1) |
| 8916 | .nr(16) |
| 8917 | .kr(4) |
| 8918 | .sr(1) |
| 8919 | .m(m) |
| 8920 | .n(n) |
| 8921 | .k(k) |
| 8922 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8923 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8924 | } |
| 8925 | } |
| 8926 | } |
| 8927 | } |
| 8928 | |
| 8929 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_div_16) { |
| 8930 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8931 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8932 | for (size_t k = 1; k <= 40; k += 9) { |
| 8933 | GemmMicrokernelTester() |
| 8934 | .mr(1) |
| 8935 | .nr(16) |
| 8936 | .kr(4) |
| 8937 | .sr(1) |
| 8938 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8939 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8940 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8941 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8942 | } |
| 8943 | } |
| 8944 | } |
| 8945 | |
| 8946 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_div_16_strided_cn) { |
| 8947 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8948 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8949 | for (size_t k = 1; k <= 40; k += 9) { |
| 8950 | GemmMicrokernelTester() |
| 8951 | .mr(1) |
| 8952 | .nr(16) |
| 8953 | .kr(4) |
| 8954 | .sr(1) |
| 8955 | .m(1) |
| 8956 | .n(n) |
| 8957 | .k(k) |
| 8958 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8959 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8960 | } |
| 8961 | } |
| 8962 | } |
| 8963 | |
| 8964 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_div_16_strided_a) { |
| 8965 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8966 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8967 | for (size_t k = 1; k <= 40; k += 9) { |
| 8968 | GemmMicrokernelTester() |
| 8969 | .mr(1) |
| 8970 | .nr(16) |
| 8971 | .kr(4) |
| 8972 | .sr(1) |
| 8973 | .m(1) |
| 8974 | .n(n) |
| 8975 | .k(k) |
| 8976 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8977 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8978 | } |
| 8979 | } |
| 8980 | } |
| 8981 | |
| 8982 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, n_div_16_subtile) { |
| 8983 | TEST_REQUIRES_ARM_NEON_DOT; |
| 8984 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8985 | for (size_t k = 1; k <= 40; k += 9) { |
| 8986 | for (uint32_t m = 1; m <= 1; m++) { |
| 8987 | GemmMicrokernelTester() |
| 8988 | .mr(1) |
| 8989 | .nr(16) |
| 8990 | .kr(4) |
| 8991 | .sr(1) |
| 8992 | .m(m) |
| 8993 | .n(n) |
| 8994 | .k(k) |
| 8995 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8996 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 8997 | } |
| 8998 | } |
| 8999 | } |
| 9000 | } |
| 9001 | |
| 9002 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, strided_cm_subtile) { |
| 9003 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9004 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9005 | for (uint32_t n = 1; n <= 16; n++) { |
| 9006 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9007 | GemmMicrokernelTester() |
| 9008 | .mr(1) |
| 9009 | .nr(16) |
| 9010 | .kr(4) |
| 9011 | .sr(1) |
| 9012 | .m(m) |
| 9013 | .n(n) |
| 9014 | .k(k) |
| 9015 | .cm_stride(19) |
| 9016 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9017 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9018 | } |
| 9019 | } |
| 9020 | } |
| 9021 | } |
| 9022 | |
| 9023 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, qmin) { |
| 9024 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9025 | GemmMicrokernelTester() |
| 9026 | .mr(1) |
| 9027 | .nr(16) |
| 9028 | .kr(4) |
| 9029 | .sr(1) |
| 9030 | .m(1) |
| 9031 | .n(16) |
| 9032 | .k(8) |
| 9033 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9034 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9035 | } |
| 9036 | |
| 9037 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, qmax) { |
| 9038 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9039 | GemmMicrokernelTester() |
| 9040 | .mr(1) |
| 9041 | .nr(16) |
| 9042 | .kr(4) |
| 9043 | .sr(1) |
| 9044 | .m(1) |
| 9045 | .n(16) |
| 9046 | .k(8) |
| 9047 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9048 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9049 | } |
| 9050 | |
| 9051 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, strided_cm) { |
| 9052 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9053 | GemmMicrokernelTester() |
| 9054 | .mr(1) |
| 9055 | .nr(16) |
| 9056 | .kr(4) |
| 9057 | .sr(1) |
| 9058 | .m(1) |
| 9059 | .n(16) |
| 9060 | .k(8) |
| 9061 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9062 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9063 | } |
| 9064 | |
| 9065 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, no_a_zero_point) { |
| 9066 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9067 | for (size_t k = 1; k <= 40; k += 9) { |
| 9068 | GemmMicrokernelTester() |
| 9069 | .mr(1) |
| 9070 | .nr(16) |
| 9071 | .kr(4) |
| 9072 | .sr(1) |
| 9073 | .m(1) |
| 9074 | .n(16) |
| 9075 | .k(k) |
| 9076 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9077 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9078 | } |
| 9079 | } |
| 9080 | |
| 9081 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, no_b_zero_point) { |
| 9082 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9083 | for (size_t k = 1; k <= 40; k += 9) { |
| 9084 | GemmMicrokernelTester() |
| 9085 | .mr(1) |
| 9086 | .nr(16) |
| 9087 | .kr(4) |
| 9088 | .sr(1) |
| 9089 | .m(1) |
| 9090 | .n(16) |
| 9091 | .k(k) |
| 9092 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9093 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9094 | } |
| 9095 | } |
| 9096 | |
| 9097 | TEST(QU8_GEMM_MINMAX_RNDNU_1X16C4__NEONDOT, no_zero_point) { |
| 9098 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9099 | for (size_t k = 1; k <= 40; k += 9) { |
| 9100 | GemmMicrokernelTester() |
| 9101 | .mr(1) |
| 9102 | .nr(16) |
| 9103 | .kr(4) |
| 9104 | .sr(1) |
| 9105 | .m(1) |
| 9106 | .n(16) |
| 9107 | .k(k) |
| 9108 | .a_zero_point(0) |
| 9109 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9110 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9111 | } |
| 9112 | } |
| 9113 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 9114 | |
| 9115 | |
| 9116 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9117 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8) { |
| 9118 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9119 | GemmMicrokernelTester() |
| 9120 | .mr(6) |
| 9121 | .nr(16) |
| 9122 | .kr(4) |
| 9123 | .sr(1) |
| 9124 | .m(6) |
| 9125 | .n(16) |
| 9126 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9127 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9128 | } |
| 9129 | |
| 9130 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cn) { |
| 9131 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9132 | GemmMicrokernelTester() |
| 9133 | .mr(6) |
| 9134 | .nr(16) |
| 9135 | .kr(4) |
| 9136 | .sr(1) |
| 9137 | .m(6) |
| 9138 | .n(16) |
| 9139 | .k(8) |
| 9140 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9141 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9142 | } |
| 9143 | |
| 9144 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_strided_a) { |
| 9145 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9146 | GemmMicrokernelTester() |
| 9147 | .mr(6) |
| 9148 | .nr(16) |
| 9149 | .kr(4) |
| 9150 | .sr(1) |
| 9151 | .m(6) |
| 9152 | .n(16) |
| 9153 | .k(8) |
| 9154 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9155 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9156 | } |
| 9157 | |
| 9158 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile) { |
| 9159 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9160 | for (uint32_t n = 1; n <= 16; n++) { |
| 9161 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9162 | GemmMicrokernelTester() |
| 9163 | .mr(6) |
| 9164 | .nr(16) |
| 9165 | .kr(4) |
| 9166 | .sr(1) |
| 9167 | .m(m) |
| 9168 | .n(n) |
| 9169 | .k(8) |
| 9170 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9171 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9172 | } |
| 9173 | } |
| 9174 | } |
| 9175 | |
| 9176 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile_m) { |
| 9177 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9178 | for (uint32_t m = 1; m <= 6; m++) { |
| 9179 | GemmMicrokernelTester() |
| 9180 | .mr(6) |
| 9181 | .nr(16) |
| 9182 | .kr(4) |
| 9183 | .sr(1) |
| 9184 | .m(m) |
| 9185 | .n(16) |
| 9186 | .k(8) |
| 9187 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9188 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9189 | } |
| 9190 | } |
| 9191 | |
| 9192 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile_n) { |
| 9193 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9194 | for (uint32_t n = 1; n <= 16; n++) { |
| 9195 | GemmMicrokernelTester() |
| 9196 | .mr(6) |
| 9197 | .nr(16) |
| 9198 | .kr(4) |
| 9199 | .sr(1) |
| 9200 | .m(6) |
| 9201 | .n(n) |
| 9202 | .k(8) |
| 9203 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9204 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9205 | } |
| 9206 | } |
| 9207 | |
| 9208 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8) { |
| 9209 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9210 | for (size_t k = 1; k < 8; k++) { |
| 9211 | GemmMicrokernelTester() |
| 9212 | .mr(6) |
| 9213 | .nr(16) |
| 9214 | .kr(4) |
| 9215 | .sr(1) |
| 9216 | .m(6) |
| 9217 | .n(16) |
| 9218 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9219 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9220 | } |
| 9221 | } |
| 9222 | |
| 9223 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8_strided_a) { |
| 9224 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9225 | for (size_t k = 1; k < 8; k++) { |
| 9226 | GemmMicrokernelTester() |
| 9227 | .mr(6) |
| 9228 | .nr(16) |
| 9229 | .kr(4) |
| 9230 | .sr(1) |
| 9231 | .m(6) |
| 9232 | .n(16) |
| 9233 | .k(k) |
| 9234 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9235 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9236 | } |
| 9237 | } |
| 9238 | |
| 9239 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8_subtile) { |
| 9240 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9241 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9242 | for (uint32_t n = 1; n <= 16; n++) { |
| 9243 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9244 | GemmMicrokernelTester() |
| 9245 | .mr(6) |
| 9246 | .nr(16) |
| 9247 | .kr(4) |
| 9248 | .sr(1) |
| 9249 | .m(m) |
| 9250 | .n(n) |
| 9251 | .k(k) |
| 9252 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9253 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9254 | } |
| 9255 | } |
| 9256 | } |
| 9257 | } |
| 9258 | |
| 9259 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8) { |
| 9260 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9261 | for (size_t k = 9; k < 16; k++) { |
| 9262 | GemmMicrokernelTester() |
| 9263 | .mr(6) |
| 9264 | .nr(16) |
| 9265 | .kr(4) |
| 9266 | .sr(1) |
| 9267 | .m(6) |
| 9268 | .n(16) |
| 9269 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9270 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9271 | } |
| 9272 | } |
| 9273 | |
| 9274 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8_strided_a) { |
| 9275 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9276 | for (size_t k = 9; k < 16; k++) { |
| 9277 | GemmMicrokernelTester() |
| 9278 | .mr(6) |
| 9279 | .nr(16) |
| 9280 | .kr(4) |
| 9281 | .sr(1) |
| 9282 | .m(6) |
| 9283 | .n(16) |
| 9284 | .k(k) |
| 9285 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9286 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9287 | } |
| 9288 | } |
| 9289 | |
| 9290 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8_subtile) { |
| 9291 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9292 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9293 | for (uint32_t n = 1; n <= 16; n++) { |
| 9294 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9295 | GemmMicrokernelTester() |
| 9296 | .mr(6) |
| 9297 | .nr(16) |
| 9298 | .kr(4) |
| 9299 | .sr(1) |
| 9300 | .m(m) |
| 9301 | .n(n) |
| 9302 | .k(k) |
| 9303 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9304 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9305 | } |
| 9306 | } |
| 9307 | } |
| 9308 | } |
| 9309 | |
| 9310 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8) { |
| 9311 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9312 | for (size_t k = 16; k <= 80; k += 8) { |
| 9313 | GemmMicrokernelTester() |
| 9314 | .mr(6) |
| 9315 | .nr(16) |
| 9316 | .kr(4) |
| 9317 | .sr(1) |
| 9318 | .m(6) |
| 9319 | .n(16) |
| 9320 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9321 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9322 | } |
| 9323 | } |
| 9324 | |
| 9325 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8_strided_a) { |
| 9326 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9327 | for (size_t k = 16; k <= 80; k += 8) { |
| 9328 | GemmMicrokernelTester() |
| 9329 | .mr(6) |
| 9330 | .nr(16) |
| 9331 | .kr(4) |
| 9332 | .sr(1) |
| 9333 | .m(6) |
| 9334 | .n(16) |
| 9335 | .k(k) |
| 9336 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9337 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9338 | } |
| 9339 | } |
| 9340 | |
| 9341 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8_subtile) { |
| 9342 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9343 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9344 | for (uint32_t n = 1; n <= 16; n++) { |
| 9345 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9346 | GemmMicrokernelTester() |
| 9347 | .mr(6) |
| 9348 | .nr(16) |
| 9349 | .kr(4) |
| 9350 | .sr(1) |
| 9351 | .m(m) |
| 9352 | .n(n) |
| 9353 | .k(k) |
| 9354 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9355 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9356 | } |
| 9357 | } |
| 9358 | } |
| 9359 | } |
| 9360 | |
| 9361 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16) { |
| 9362 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9363 | for (uint32_t n = 17; n < 32; n++) { |
| 9364 | for (size_t k = 1; k <= 40; k += 9) { |
| 9365 | GemmMicrokernelTester() |
| 9366 | .mr(6) |
| 9367 | .nr(16) |
| 9368 | .kr(4) |
| 9369 | .sr(1) |
| 9370 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9371 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9372 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9373 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9374 | } |
| 9375 | } |
| 9376 | } |
| 9377 | |
| 9378 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_strided_cn) { |
| 9379 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9380 | for (uint32_t n = 17; n < 32; n++) { |
| 9381 | for (size_t k = 1; k <= 40; k += 9) { |
| 9382 | GemmMicrokernelTester() |
| 9383 | .mr(6) |
| 9384 | .nr(16) |
| 9385 | .kr(4) |
| 9386 | .sr(1) |
| 9387 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9388 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9389 | .k(k) |
| 9390 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9391 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9392 | } |
| 9393 | } |
| 9394 | } |
| 9395 | |
| 9396 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_strided_a) { |
| 9397 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9398 | for (uint32_t n = 17; n < 32; n++) { |
| 9399 | for (size_t k = 1; k <= 40; k += 9) { |
| 9400 | GemmMicrokernelTester() |
| 9401 | .mr(6) |
| 9402 | .nr(16) |
| 9403 | .kr(4) |
| 9404 | .sr(1) |
| 9405 | .m(6) |
| 9406 | .n(n) |
| 9407 | .k(k) |
| 9408 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9409 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9410 | } |
| 9411 | } |
| 9412 | } |
| 9413 | |
| 9414 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_subtile) { |
| 9415 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9416 | for (uint32_t n = 17; n < 32; n++) { |
| 9417 | for (size_t k = 1; k <= 40; k += 9) { |
| 9418 | for (uint32_t m = 1; m <= 6; m++) { |
| 9419 | GemmMicrokernelTester() |
| 9420 | .mr(6) |
| 9421 | .nr(16) |
| 9422 | .kr(4) |
| 9423 | .sr(1) |
| 9424 | .m(m) |
| 9425 | .n(n) |
| 9426 | .k(k) |
| 9427 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9428 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9429 | } |
| 9430 | } |
| 9431 | } |
| 9432 | } |
| 9433 | |
| 9434 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16) { |
| 9435 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9436 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9437 | for (size_t k = 1; k <= 40; k += 9) { |
| 9438 | GemmMicrokernelTester() |
| 9439 | .mr(6) |
| 9440 | .nr(16) |
| 9441 | .kr(4) |
| 9442 | .sr(1) |
| 9443 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9444 | .n(n) |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9445 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9446 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9447 | } |
| 9448 | } |
| 9449 | } |
| 9450 | |
| 9451 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_strided_cn) { |
| 9452 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9453 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9454 | for (size_t k = 1; k <= 40; k += 9) { |
| 9455 | GemmMicrokernelTester() |
| 9456 | .mr(6) |
| 9457 | .nr(16) |
| 9458 | .kr(4) |
| 9459 | .sr(1) |
| 9460 | .m(6) |
| 9461 | .n(n) |
| 9462 | .k(k) |
| 9463 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9464 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9465 | } |
| 9466 | } |
| 9467 | } |
| 9468 | |
| 9469 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_strided_a) { |
| 9470 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9471 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9472 | for (size_t k = 1; k <= 40; k += 9) { |
| 9473 | GemmMicrokernelTester() |
| 9474 | .mr(6) |
| 9475 | .nr(16) |
| 9476 | .kr(4) |
| 9477 | .sr(1) |
| 9478 | .m(6) |
| 9479 | .n(n) |
| 9480 | .k(k) |
| 9481 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9482 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9483 | } |
| 9484 | } |
| 9485 | } |
| 9486 | |
| 9487 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_subtile) { |
| 9488 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9489 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9490 | for (size_t k = 1; k <= 40; k += 9) { |
| 9491 | for (uint32_t m = 1; m <= 6; m++) { |
| 9492 | GemmMicrokernelTester() |
| 9493 | .mr(6) |
| 9494 | .nr(16) |
| 9495 | .kr(4) |
| 9496 | .sr(1) |
| 9497 | .m(m) |
| 9498 | .n(n) |
| 9499 | .k(k) |
| 9500 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9501 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9502 | } |
| 9503 | } |
| 9504 | } |
| 9505 | } |
| 9506 | |
| 9507 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cm_subtile) { |
| 9508 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9509 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9510 | for (uint32_t n = 1; n <= 16; n++) { |
| 9511 | for (uint32_t m = 1; m <= 6; m++) { |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9512 | GemmMicrokernelTester() |
| 9513 | .mr(6) |
| 9514 | .nr(16) |
| 9515 | .kr(4) |
| 9516 | .sr(1) |
| 9517 | .m(m) |
| 9518 | .n(n) |
| 9519 | .k(k) |
| 9520 | .cm_stride(19) |
| 9521 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9522 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9523 | } |
| 9524 | } |
| 9525 | } |
| 9526 | } |
| 9527 | |
| 9528 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, qmin) { |
| 9529 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9530 | GemmMicrokernelTester() |
| 9531 | .mr(6) |
| 9532 | .nr(16) |
| 9533 | .kr(4) |
| 9534 | .sr(1) |
| 9535 | .m(6) |
| 9536 | .n(16) |
| 9537 | .k(8) |
| 9538 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9539 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9540 | } |
| 9541 | |
| 9542 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, qmax) { |
| 9543 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9544 | GemmMicrokernelTester() |
| 9545 | .mr(6) |
| 9546 | .nr(16) |
| 9547 | .kr(4) |
| 9548 | .sr(1) |
| 9549 | .m(6) |
| 9550 | .n(16) |
| 9551 | .k(8) |
| 9552 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9553 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9554 | } |
| 9555 | |
| 9556 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cm) { |
| 9557 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9558 | GemmMicrokernelTester() |
| 9559 | .mr(6) |
| 9560 | .nr(16) |
| 9561 | .kr(4) |
| 9562 | .sr(1) |
| 9563 | .m(6) |
| 9564 | .n(16) |
| 9565 | .k(8) |
| 9566 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9567 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9568 | } |
| 9569 | |
| 9570 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, no_a_zero_point) { |
| 9571 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9572 | for (size_t k = 1; k <= 40; k += 9) { |
| 9573 | GemmMicrokernelTester() |
| 9574 | .mr(6) |
| 9575 | .nr(16) |
| 9576 | .kr(4) |
| 9577 | .sr(1) |
| 9578 | .m(6) |
| 9579 | .n(16) |
| 9580 | .k(k) |
| 9581 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9582 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9583 | } |
| 9584 | } |
| 9585 | |
| 9586 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, no_b_zero_point) { |
| 9587 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9588 | for (size_t k = 1; k <= 40; k += 9) { |
| 9589 | GemmMicrokernelTester() |
| 9590 | .mr(6) |
| 9591 | .nr(16) |
| 9592 | .kr(4) |
| 9593 | .sr(1) |
| 9594 | .m(6) |
| 9595 | .n(16) |
| 9596 | .k(k) |
| 9597 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9598 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9599 | } |
| 9600 | } |
| 9601 | |
| 9602 | TEST(QU8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, no_zero_point) { |
| 9603 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9604 | for (size_t k = 1; k <= 40; k += 9) { |
| 9605 | GemmMicrokernelTester() |
| 9606 | .mr(6) |
| 9607 | .nr(16) |
| 9608 | .kr(4) |
| 9609 | .sr(1) |
| 9610 | .m(6) |
| 9611 | .n(16) |
| 9612 | .k(k) |
| 9613 | .a_zero_point(0) |
| 9614 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9615 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 88e839c | 2021-08-11 00:12:31 -0700 | [diff] [blame] | 9616 | } |
| 9617 | } |
| 9618 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 9619 | |
| 9620 | |
| 9621 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9622 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_eq_8) { |
| 9623 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9624 | GemmMicrokernelTester() |
| 9625 | .mr(1) |
| 9626 | .nr(32) |
| 9627 | .kr(4) |
| 9628 | .sr(1) |
| 9629 | .m(1) |
| 9630 | .n(32) |
| 9631 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9632 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9633 | } |
| 9634 | |
| 9635 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, strided_cn) { |
| 9636 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9637 | GemmMicrokernelTester() |
| 9638 | .mr(1) |
| 9639 | .nr(32) |
| 9640 | .kr(4) |
| 9641 | .sr(1) |
| 9642 | .m(1) |
| 9643 | .n(32) |
| 9644 | .k(8) |
| 9645 | .cn_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9646 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9647 | } |
| 9648 | |
| 9649 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_eq_8_strided_a) { |
| 9650 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9651 | GemmMicrokernelTester() |
| 9652 | .mr(1) |
| 9653 | .nr(32) |
| 9654 | .kr(4) |
| 9655 | .sr(1) |
| 9656 | .m(1) |
| 9657 | .n(32) |
| 9658 | .k(8) |
| 9659 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9660 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9661 | } |
| 9662 | |
| 9663 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_eq_8_subtile) { |
| 9664 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9665 | for (uint32_t n = 1; n <= 32; n++) { |
| 9666 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9667 | GemmMicrokernelTester() |
| 9668 | .mr(1) |
| 9669 | .nr(32) |
| 9670 | .kr(4) |
| 9671 | .sr(1) |
| 9672 | .m(m) |
| 9673 | .n(n) |
| 9674 | .k(8) |
| 9675 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9676 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9677 | } |
| 9678 | } |
| 9679 | } |
| 9680 | |
| 9681 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_eq_8_subtile_m) { |
| 9682 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9683 | for (uint32_t m = 1; m <= 1; m++) { |
| 9684 | GemmMicrokernelTester() |
| 9685 | .mr(1) |
| 9686 | .nr(32) |
| 9687 | .kr(4) |
| 9688 | .sr(1) |
| 9689 | .m(m) |
| 9690 | .n(32) |
| 9691 | .k(8) |
| 9692 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9693 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9694 | } |
| 9695 | } |
| 9696 | |
| 9697 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_eq_8_subtile_n) { |
| 9698 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9699 | for (uint32_t n = 1; n <= 32; n++) { |
| 9700 | GemmMicrokernelTester() |
| 9701 | .mr(1) |
| 9702 | .nr(32) |
| 9703 | .kr(4) |
| 9704 | .sr(1) |
| 9705 | .m(1) |
| 9706 | .n(n) |
| 9707 | .k(8) |
| 9708 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9709 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9710 | } |
| 9711 | } |
| 9712 | |
| 9713 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_lt_8) { |
| 9714 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9715 | for (size_t k = 1; k < 8; k++) { |
| 9716 | GemmMicrokernelTester() |
| 9717 | .mr(1) |
| 9718 | .nr(32) |
| 9719 | .kr(4) |
| 9720 | .sr(1) |
| 9721 | .m(1) |
| 9722 | .n(32) |
| 9723 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9724 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9725 | } |
| 9726 | } |
| 9727 | |
| 9728 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_lt_8_strided_a) { |
| 9729 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9730 | for (size_t k = 1; k < 8; k++) { |
| 9731 | GemmMicrokernelTester() |
| 9732 | .mr(1) |
| 9733 | .nr(32) |
| 9734 | .kr(4) |
| 9735 | .sr(1) |
| 9736 | .m(1) |
| 9737 | .n(32) |
| 9738 | .k(k) |
| 9739 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9740 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9741 | } |
| 9742 | } |
| 9743 | |
| 9744 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_lt_8_subtile) { |
| 9745 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9746 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9747 | for (uint32_t n = 1; n <= 32; n++) { |
| 9748 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9749 | GemmMicrokernelTester() |
| 9750 | .mr(1) |
| 9751 | .nr(32) |
| 9752 | .kr(4) |
| 9753 | .sr(1) |
| 9754 | .m(m) |
| 9755 | .n(n) |
| 9756 | .k(k) |
| 9757 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9758 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9759 | } |
| 9760 | } |
| 9761 | } |
| 9762 | } |
| 9763 | |
| 9764 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_gt_8) { |
| 9765 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9766 | for (size_t k = 9; k < 16; k++) { |
| 9767 | GemmMicrokernelTester() |
| 9768 | .mr(1) |
| 9769 | .nr(32) |
| 9770 | .kr(4) |
| 9771 | .sr(1) |
| 9772 | .m(1) |
| 9773 | .n(32) |
| 9774 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9775 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9776 | } |
| 9777 | } |
| 9778 | |
| 9779 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_gt_8_strided_a) { |
| 9780 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9781 | for (size_t k = 9; k < 16; k++) { |
| 9782 | GemmMicrokernelTester() |
| 9783 | .mr(1) |
| 9784 | .nr(32) |
| 9785 | .kr(4) |
| 9786 | .sr(1) |
| 9787 | .m(1) |
| 9788 | .n(32) |
| 9789 | .k(k) |
| 9790 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9791 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9792 | } |
| 9793 | } |
| 9794 | |
| 9795 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_gt_8_subtile) { |
| 9796 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9797 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9798 | for (uint32_t n = 1; n <= 32; n++) { |
| 9799 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9800 | GemmMicrokernelTester() |
| 9801 | .mr(1) |
| 9802 | .nr(32) |
| 9803 | .kr(4) |
| 9804 | .sr(1) |
| 9805 | .m(m) |
| 9806 | .n(n) |
| 9807 | .k(k) |
| 9808 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9809 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9810 | } |
| 9811 | } |
| 9812 | } |
| 9813 | } |
| 9814 | |
| 9815 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_div_8) { |
| 9816 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9817 | for (size_t k = 16; k <= 80; k += 8) { |
| 9818 | GemmMicrokernelTester() |
| 9819 | .mr(1) |
| 9820 | .nr(32) |
| 9821 | .kr(4) |
| 9822 | .sr(1) |
| 9823 | .m(1) |
| 9824 | .n(32) |
| 9825 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9826 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9827 | } |
| 9828 | } |
| 9829 | |
| 9830 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_div_8_strided_a) { |
| 9831 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9832 | for (size_t k = 16; k <= 80; k += 8) { |
| 9833 | GemmMicrokernelTester() |
| 9834 | .mr(1) |
| 9835 | .nr(32) |
| 9836 | .kr(4) |
| 9837 | .sr(1) |
| 9838 | .m(1) |
| 9839 | .n(32) |
| 9840 | .k(k) |
| 9841 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9842 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9843 | } |
| 9844 | } |
| 9845 | |
| 9846 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, k_div_8_subtile) { |
| 9847 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9848 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9849 | for (uint32_t n = 1; n <= 32; n++) { |
| 9850 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9851 | GemmMicrokernelTester() |
| 9852 | .mr(1) |
| 9853 | .nr(32) |
| 9854 | .kr(4) |
| 9855 | .sr(1) |
| 9856 | .m(m) |
| 9857 | .n(n) |
| 9858 | .k(k) |
| 9859 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9860 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9861 | } |
| 9862 | } |
| 9863 | } |
| 9864 | } |
| 9865 | |
| 9866 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_gt_32) { |
| 9867 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9868 | for (uint32_t n = 33; n < 64; n++) { |
| 9869 | for (size_t k = 1; k <= 40; k += 9) { |
| 9870 | GemmMicrokernelTester() |
| 9871 | .mr(1) |
| 9872 | .nr(32) |
| 9873 | .kr(4) |
| 9874 | .sr(1) |
| 9875 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9876 | .n(n) |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9877 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9878 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9879 | } |
| 9880 | } |
| 9881 | } |
| 9882 | |
| 9883 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_gt_32_strided_cn) { |
| 9884 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9885 | for (uint32_t n = 33; n < 64; n++) { |
| 9886 | for (size_t k = 1; k <= 40; k += 9) { |
| 9887 | GemmMicrokernelTester() |
| 9888 | .mr(1) |
| 9889 | .nr(32) |
| 9890 | .kr(4) |
| 9891 | .sr(1) |
| 9892 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9893 | .n(n) |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9894 | .k(k) |
| 9895 | .cn_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9896 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9897 | } |
| 9898 | } |
| 9899 | } |
| 9900 | |
| 9901 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_gt_32_strided_a) { |
| 9902 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9903 | for (uint32_t n = 33; n < 64; n++) { |
| 9904 | for (size_t k = 1; k <= 40; k += 9) { |
| 9905 | GemmMicrokernelTester() |
| 9906 | .mr(1) |
| 9907 | .nr(32) |
| 9908 | .kr(4) |
| 9909 | .sr(1) |
| 9910 | .m(1) |
| 9911 | .n(n) |
| 9912 | .k(k) |
| 9913 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9914 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9915 | } |
| 9916 | } |
| 9917 | } |
| 9918 | |
| 9919 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_gt_32_subtile) { |
| 9920 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9921 | for (uint32_t n = 33; n < 64; n++) { |
| 9922 | for (size_t k = 1; k <= 40; k += 9) { |
| 9923 | for (uint32_t m = 1; m <= 1; m++) { |
| 9924 | GemmMicrokernelTester() |
| 9925 | .mr(1) |
| 9926 | .nr(32) |
| 9927 | .kr(4) |
| 9928 | .sr(1) |
| 9929 | .m(m) |
| 9930 | .n(n) |
| 9931 | .k(k) |
| 9932 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9933 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9934 | } |
| 9935 | } |
| 9936 | } |
| 9937 | } |
| 9938 | |
| 9939 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_div_32) { |
| 9940 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9941 | for (uint32_t n = 64; n <= 96; n += 32) { |
| 9942 | for (size_t k = 1; k <= 40; k += 9) { |
| 9943 | GemmMicrokernelTester() |
| 9944 | .mr(1) |
| 9945 | .nr(32) |
| 9946 | .kr(4) |
| 9947 | .sr(1) |
| 9948 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9949 | .n(n) |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9950 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9951 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9952 | } |
| 9953 | } |
| 9954 | } |
| 9955 | |
| 9956 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_div_32_strided_cn) { |
| 9957 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9958 | for (uint32_t n = 64; n <= 96; n += 32) { |
| 9959 | for (size_t k = 1; k <= 40; k += 9) { |
| 9960 | GemmMicrokernelTester() |
| 9961 | .mr(1) |
| 9962 | .nr(32) |
| 9963 | .kr(4) |
| 9964 | .sr(1) |
| 9965 | .m(1) |
| 9966 | .n(n) |
| 9967 | .k(k) |
| 9968 | .cn_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9969 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9970 | } |
| 9971 | } |
| 9972 | } |
| 9973 | |
| 9974 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_div_32_strided_a) { |
| 9975 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9976 | for (uint32_t n = 64; n <= 96; n += 32) { |
| 9977 | for (size_t k = 1; k <= 40; k += 9) { |
| 9978 | GemmMicrokernelTester() |
| 9979 | .mr(1) |
| 9980 | .nr(32) |
| 9981 | .kr(4) |
| 9982 | .sr(1) |
| 9983 | .m(1) |
| 9984 | .n(n) |
| 9985 | .k(k) |
| 9986 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9987 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 9988 | } |
| 9989 | } |
| 9990 | } |
| 9991 | |
| 9992 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, n_div_32_subtile) { |
| 9993 | TEST_REQUIRES_ARM_NEON_DOT; |
| 9994 | for (uint32_t n = 64; n <= 96; n += 32) { |
| 9995 | for (size_t k = 1; k <= 40; k += 9) { |
| 9996 | for (uint32_t m = 1; m <= 1; m++) { |
| 9997 | GemmMicrokernelTester() |
| 9998 | .mr(1) |
| 9999 | .nr(32) |
| 10000 | .kr(4) |
| 10001 | .sr(1) |
| 10002 | .m(m) |
| 10003 | .n(n) |
| 10004 | .k(k) |
| 10005 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10006 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10007 | } |
| 10008 | } |
| 10009 | } |
| 10010 | } |
| 10011 | |
| 10012 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, strided_cm_subtile) { |
| 10013 | TEST_REQUIRES_ARM_NEON_DOT; |
| 10014 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10015 | for (uint32_t n = 1; n <= 32; n++) { |
| 10016 | for (uint32_t m = 1; m <= 1; m++) { |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10017 | GemmMicrokernelTester() |
| 10018 | .mr(1) |
| 10019 | .nr(32) |
| 10020 | .kr(4) |
| 10021 | .sr(1) |
| 10022 | .m(m) |
| 10023 | .n(n) |
| 10024 | .k(k) |
| 10025 | .cm_stride(37) |
| 10026 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10027 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10028 | } |
| 10029 | } |
| 10030 | } |
| 10031 | } |
| 10032 | |
| 10033 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, qmin) { |
| 10034 | TEST_REQUIRES_ARM_NEON_DOT; |
| 10035 | GemmMicrokernelTester() |
| 10036 | .mr(1) |
| 10037 | .nr(32) |
| 10038 | .kr(4) |
| 10039 | .sr(1) |
| 10040 | .m(1) |
| 10041 | .n(32) |
| 10042 | .k(8) |
| 10043 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10044 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10045 | } |
| 10046 | |
| 10047 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, qmax) { |
| 10048 | TEST_REQUIRES_ARM_NEON_DOT; |
| 10049 | GemmMicrokernelTester() |
| 10050 | .mr(1) |
| 10051 | .nr(32) |
| 10052 | .kr(4) |
| 10053 | .sr(1) |
| 10054 | .m(1) |
| 10055 | .n(32) |
| 10056 | .k(8) |
| 10057 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10058 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10059 | } |
| 10060 | |
| 10061 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, strided_cm) { |
| 10062 | TEST_REQUIRES_ARM_NEON_DOT; |
| 10063 | GemmMicrokernelTester() |
| 10064 | .mr(1) |
| 10065 | .nr(32) |
| 10066 | .kr(4) |
| 10067 | .sr(1) |
| 10068 | .m(1) |
| 10069 | .n(32) |
| 10070 | .k(8) |
| 10071 | .cm_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10072 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10073 | } |
| 10074 | |
| 10075 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, no_a_zero_point) { |
| 10076 | TEST_REQUIRES_ARM_NEON_DOT; |
| 10077 | for (size_t k = 1; k <= 40; k += 9) { |
| 10078 | GemmMicrokernelTester() |
| 10079 | .mr(1) |
| 10080 | .nr(32) |
| 10081 | .kr(4) |
| 10082 | .sr(1) |
| 10083 | .m(1) |
| 10084 | .n(32) |
| 10085 | .k(k) |
| 10086 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10087 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10088 | } |
| 10089 | } |
| 10090 | |
| 10091 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, no_b_zero_point) { |
| 10092 | TEST_REQUIRES_ARM_NEON_DOT; |
| 10093 | for (size_t k = 1; k <= 40; k += 9) { |
| 10094 | GemmMicrokernelTester() |
| 10095 | .mr(1) |
| 10096 | .nr(32) |
| 10097 | .kr(4) |
| 10098 | .sr(1) |
| 10099 | .m(1) |
| 10100 | .n(32) |
| 10101 | .k(k) |
| 10102 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10103 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10104 | } |
| 10105 | } |
| 10106 | |
| 10107 | TEST(QU8_GEMM_MINMAX_RNDNU_1X32C4__NEONDOT, no_zero_point) { |
| 10108 | TEST_REQUIRES_ARM_NEON_DOT; |
| 10109 | for (size_t k = 1; k <= 40; k += 9) { |
| 10110 | GemmMicrokernelTester() |
| 10111 | .mr(1) |
| 10112 | .nr(32) |
| 10113 | .kr(4) |
| 10114 | .sr(1) |
| 10115 | .m(1) |
| 10116 | .n(32) |
| 10117 | .k(k) |
| 10118 | .a_zero_point(0) |
| 10119 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10120 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_1x32c4__neondot, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | cdf59a5 | 2021-09-08 13:55:24 -0700 | [diff] [blame] | 10121 | } |
| 10122 | } |
| 10123 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 10124 | |
| 10125 | |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 10126 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10127 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_eq_8) { |
| 10128 | TEST_REQUIRES_ARM_NEON; |
| 10129 | GemmMicrokernelTester() |
| 10130 | .mr(4) |
| 10131 | .nr(16) |
| 10132 | .kr(1) |
| 10133 | .sr(1) |
| 10134 | .m(4) |
| 10135 | .n(16) |
| 10136 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10137 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10138 | } |
| 10139 | |
| 10140 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, strided_cn) { |
| 10141 | TEST_REQUIRES_ARM_NEON; |
| 10142 | GemmMicrokernelTester() |
| 10143 | .mr(4) |
| 10144 | .nr(16) |
| 10145 | .kr(1) |
| 10146 | .sr(1) |
| 10147 | .m(4) |
| 10148 | .n(16) |
| 10149 | .k(8) |
| 10150 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10151 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10152 | } |
| 10153 | |
| 10154 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_eq_8_strided_a) { |
| 10155 | TEST_REQUIRES_ARM_NEON; |
| 10156 | GemmMicrokernelTester() |
| 10157 | .mr(4) |
| 10158 | .nr(16) |
| 10159 | .kr(1) |
| 10160 | .sr(1) |
| 10161 | .m(4) |
| 10162 | .n(16) |
| 10163 | .k(8) |
| 10164 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10165 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10166 | } |
| 10167 | |
| 10168 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_eq_8_subtile) { |
| 10169 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10170 | for (uint32_t n = 1; n <= 16; n++) { |
| 10171 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10172 | GemmMicrokernelTester() |
| 10173 | .mr(4) |
| 10174 | .nr(16) |
| 10175 | .kr(1) |
| 10176 | .sr(1) |
| 10177 | .m(m) |
| 10178 | .n(n) |
| 10179 | .k(8) |
| 10180 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10181 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10182 | } |
| 10183 | } |
| 10184 | } |
| 10185 | |
| 10186 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_eq_8_subtile_m) { |
| 10187 | TEST_REQUIRES_ARM_NEON; |
| 10188 | for (uint32_t m = 1; m <= 4; m++) { |
| 10189 | GemmMicrokernelTester() |
| 10190 | .mr(4) |
| 10191 | .nr(16) |
| 10192 | .kr(1) |
| 10193 | .sr(1) |
| 10194 | .m(m) |
| 10195 | .n(16) |
| 10196 | .k(8) |
| 10197 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10198 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10199 | } |
| 10200 | } |
| 10201 | |
| 10202 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_eq_8_subtile_n) { |
| 10203 | TEST_REQUIRES_ARM_NEON; |
| 10204 | for (uint32_t n = 1; n <= 16; n++) { |
| 10205 | GemmMicrokernelTester() |
| 10206 | .mr(4) |
| 10207 | .nr(16) |
| 10208 | .kr(1) |
| 10209 | .sr(1) |
| 10210 | .m(4) |
| 10211 | .n(n) |
| 10212 | .k(8) |
| 10213 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10214 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10215 | } |
| 10216 | } |
| 10217 | |
| 10218 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_lt_8) { |
| 10219 | TEST_REQUIRES_ARM_NEON; |
| 10220 | for (size_t k = 1; k < 8; k++) { |
| 10221 | GemmMicrokernelTester() |
| 10222 | .mr(4) |
| 10223 | .nr(16) |
| 10224 | .kr(1) |
| 10225 | .sr(1) |
| 10226 | .m(4) |
| 10227 | .n(16) |
| 10228 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10229 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10230 | } |
| 10231 | } |
| 10232 | |
| 10233 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_lt_8_strided_a) { |
| 10234 | TEST_REQUIRES_ARM_NEON; |
| 10235 | for (size_t k = 1; k < 8; k++) { |
| 10236 | GemmMicrokernelTester() |
| 10237 | .mr(4) |
| 10238 | .nr(16) |
| 10239 | .kr(1) |
| 10240 | .sr(1) |
| 10241 | .m(4) |
| 10242 | .n(16) |
| 10243 | .k(k) |
| 10244 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10245 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10246 | } |
| 10247 | } |
| 10248 | |
| 10249 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_lt_8_subtile) { |
| 10250 | TEST_REQUIRES_ARM_NEON; |
| 10251 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10252 | for (uint32_t n = 1; n <= 16; n++) { |
| 10253 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10254 | GemmMicrokernelTester() |
| 10255 | .mr(4) |
| 10256 | .nr(16) |
| 10257 | .kr(1) |
| 10258 | .sr(1) |
| 10259 | .m(m) |
| 10260 | .n(n) |
| 10261 | .k(k) |
| 10262 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10263 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10264 | } |
| 10265 | } |
| 10266 | } |
| 10267 | } |
| 10268 | |
| 10269 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_gt_8) { |
| 10270 | TEST_REQUIRES_ARM_NEON; |
| 10271 | for (size_t k = 9; k < 16; k++) { |
| 10272 | GemmMicrokernelTester() |
| 10273 | .mr(4) |
| 10274 | .nr(16) |
| 10275 | .kr(1) |
| 10276 | .sr(1) |
| 10277 | .m(4) |
| 10278 | .n(16) |
| 10279 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10280 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10281 | } |
| 10282 | } |
| 10283 | |
| 10284 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_gt_8_strided_a) { |
| 10285 | TEST_REQUIRES_ARM_NEON; |
| 10286 | for (size_t k = 9; k < 16; k++) { |
| 10287 | GemmMicrokernelTester() |
| 10288 | .mr(4) |
| 10289 | .nr(16) |
| 10290 | .kr(1) |
| 10291 | .sr(1) |
| 10292 | .m(4) |
| 10293 | .n(16) |
| 10294 | .k(k) |
| 10295 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10296 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10297 | } |
| 10298 | } |
| 10299 | |
| 10300 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_gt_8_subtile) { |
| 10301 | TEST_REQUIRES_ARM_NEON; |
| 10302 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10303 | for (uint32_t n = 1; n <= 16; n++) { |
| 10304 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10305 | GemmMicrokernelTester() |
| 10306 | .mr(4) |
| 10307 | .nr(16) |
| 10308 | .kr(1) |
| 10309 | .sr(1) |
| 10310 | .m(m) |
| 10311 | .n(n) |
| 10312 | .k(k) |
| 10313 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10314 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10315 | } |
| 10316 | } |
| 10317 | } |
| 10318 | } |
| 10319 | |
| 10320 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_div_8) { |
| 10321 | TEST_REQUIRES_ARM_NEON; |
| 10322 | for (size_t k = 16; k <= 80; k += 8) { |
| 10323 | GemmMicrokernelTester() |
| 10324 | .mr(4) |
| 10325 | .nr(16) |
| 10326 | .kr(1) |
| 10327 | .sr(1) |
| 10328 | .m(4) |
| 10329 | .n(16) |
| 10330 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10331 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10332 | } |
| 10333 | } |
| 10334 | |
| 10335 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_div_8_strided_a) { |
| 10336 | TEST_REQUIRES_ARM_NEON; |
| 10337 | for (size_t k = 16; k <= 80; k += 8) { |
| 10338 | GemmMicrokernelTester() |
| 10339 | .mr(4) |
| 10340 | .nr(16) |
| 10341 | .kr(1) |
| 10342 | .sr(1) |
| 10343 | .m(4) |
| 10344 | .n(16) |
| 10345 | .k(k) |
| 10346 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10347 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10348 | } |
| 10349 | } |
| 10350 | |
| 10351 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, k_div_8_subtile) { |
| 10352 | TEST_REQUIRES_ARM_NEON; |
| 10353 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10354 | for (uint32_t n = 1; n <= 16; n++) { |
| 10355 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10356 | GemmMicrokernelTester() |
| 10357 | .mr(4) |
| 10358 | .nr(16) |
| 10359 | .kr(1) |
| 10360 | .sr(1) |
| 10361 | .m(m) |
| 10362 | .n(n) |
| 10363 | .k(k) |
| 10364 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10365 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10366 | } |
| 10367 | } |
| 10368 | } |
| 10369 | } |
| 10370 | |
| 10371 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_gt_16) { |
| 10372 | TEST_REQUIRES_ARM_NEON; |
| 10373 | for (uint32_t n = 17; n < 32; n++) { |
| 10374 | for (size_t k = 1; k <= 40; k += 9) { |
| 10375 | GemmMicrokernelTester() |
| 10376 | .mr(4) |
| 10377 | .nr(16) |
| 10378 | .kr(1) |
| 10379 | .sr(1) |
| 10380 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10381 | .n(n) |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10382 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10383 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10384 | } |
| 10385 | } |
| 10386 | } |
| 10387 | |
| 10388 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_gt_16_strided_cn) { |
| 10389 | TEST_REQUIRES_ARM_NEON; |
| 10390 | for (uint32_t n = 17; n < 32; n++) { |
| 10391 | for (size_t k = 1; k <= 40; k += 9) { |
| 10392 | GemmMicrokernelTester() |
| 10393 | .mr(4) |
| 10394 | .nr(16) |
| 10395 | .kr(1) |
| 10396 | .sr(1) |
| 10397 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10398 | .n(n) |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10399 | .k(k) |
| 10400 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10401 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10402 | } |
| 10403 | } |
| 10404 | } |
| 10405 | |
| 10406 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_gt_16_strided_a) { |
| 10407 | TEST_REQUIRES_ARM_NEON; |
| 10408 | for (uint32_t n = 17; n < 32; n++) { |
| 10409 | for (size_t k = 1; k <= 40; k += 9) { |
| 10410 | GemmMicrokernelTester() |
| 10411 | .mr(4) |
| 10412 | .nr(16) |
| 10413 | .kr(1) |
| 10414 | .sr(1) |
| 10415 | .m(4) |
| 10416 | .n(n) |
| 10417 | .k(k) |
| 10418 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10419 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10420 | } |
| 10421 | } |
| 10422 | } |
| 10423 | |
| 10424 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_gt_16_subtile) { |
| 10425 | TEST_REQUIRES_ARM_NEON; |
| 10426 | for (uint32_t n = 17; n < 32; n++) { |
| 10427 | for (size_t k = 1; k <= 40; k += 9) { |
| 10428 | for (uint32_t m = 1; m <= 4; m++) { |
| 10429 | GemmMicrokernelTester() |
| 10430 | .mr(4) |
| 10431 | .nr(16) |
| 10432 | .kr(1) |
| 10433 | .sr(1) |
| 10434 | .m(m) |
| 10435 | .n(n) |
| 10436 | .k(k) |
| 10437 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10438 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10439 | } |
| 10440 | } |
| 10441 | } |
| 10442 | } |
| 10443 | |
| 10444 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_div_16) { |
| 10445 | TEST_REQUIRES_ARM_NEON; |
| 10446 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10447 | for (size_t k = 1; k <= 40; k += 9) { |
| 10448 | GemmMicrokernelTester() |
| 10449 | .mr(4) |
| 10450 | .nr(16) |
| 10451 | .kr(1) |
| 10452 | .sr(1) |
| 10453 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10454 | .n(n) |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10455 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10456 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10457 | } |
| 10458 | } |
| 10459 | } |
| 10460 | |
| 10461 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_div_16_strided_cn) { |
| 10462 | TEST_REQUIRES_ARM_NEON; |
| 10463 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10464 | for (size_t k = 1; k <= 40; k += 9) { |
| 10465 | GemmMicrokernelTester() |
| 10466 | .mr(4) |
| 10467 | .nr(16) |
| 10468 | .kr(1) |
| 10469 | .sr(1) |
| 10470 | .m(4) |
| 10471 | .n(n) |
| 10472 | .k(k) |
| 10473 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10474 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10475 | } |
| 10476 | } |
| 10477 | } |
| 10478 | |
| 10479 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_div_16_strided_a) { |
| 10480 | TEST_REQUIRES_ARM_NEON; |
| 10481 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10482 | for (size_t k = 1; k <= 40; k += 9) { |
| 10483 | GemmMicrokernelTester() |
| 10484 | .mr(4) |
| 10485 | .nr(16) |
| 10486 | .kr(1) |
| 10487 | .sr(1) |
| 10488 | .m(4) |
| 10489 | .n(n) |
| 10490 | .k(k) |
| 10491 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10492 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10493 | } |
| 10494 | } |
| 10495 | } |
| 10496 | |
| 10497 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, n_div_16_subtile) { |
| 10498 | TEST_REQUIRES_ARM_NEON; |
| 10499 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10500 | for (size_t k = 1; k <= 40; k += 9) { |
| 10501 | for (uint32_t m = 1; m <= 4; m++) { |
| 10502 | GemmMicrokernelTester() |
| 10503 | .mr(4) |
| 10504 | .nr(16) |
| 10505 | .kr(1) |
| 10506 | .sr(1) |
| 10507 | .m(m) |
| 10508 | .n(n) |
| 10509 | .k(k) |
| 10510 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10511 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10512 | } |
| 10513 | } |
| 10514 | } |
| 10515 | } |
| 10516 | |
| 10517 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, strided_cm_subtile) { |
| 10518 | TEST_REQUIRES_ARM_NEON; |
| 10519 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10520 | for (uint32_t n = 1; n <= 16; n++) { |
| 10521 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10522 | GemmMicrokernelTester() |
| 10523 | .mr(4) |
| 10524 | .nr(16) |
| 10525 | .kr(1) |
| 10526 | .sr(1) |
| 10527 | .m(m) |
| 10528 | .n(n) |
| 10529 | .k(k) |
| 10530 | .cm_stride(19) |
| 10531 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10532 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10533 | } |
| 10534 | } |
| 10535 | } |
| 10536 | } |
| 10537 | |
| 10538 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, qmin) { |
| 10539 | TEST_REQUIRES_ARM_NEON; |
| 10540 | GemmMicrokernelTester() |
| 10541 | .mr(4) |
| 10542 | .nr(16) |
| 10543 | .kr(1) |
| 10544 | .sr(1) |
| 10545 | .m(4) |
| 10546 | .n(16) |
| 10547 | .k(8) |
| 10548 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10549 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10550 | } |
| 10551 | |
| 10552 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, qmax) { |
| 10553 | TEST_REQUIRES_ARM_NEON; |
| 10554 | GemmMicrokernelTester() |
| 10555 | .mr(4) |
| 10556 | .nr(16) |
| 10557 | .kr(1) |
| 10558 | .sr(1) |
| 10559 | .m(4) |
| 10560 | .n(16) |
| 10561 | .k(8) |
| 10562 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10563 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10564 | } |
| 10565 | |
| 10566 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, strided_cm) { |
| 10567 | TEST_REQUIRES_ARM_NEON; |
| 10568 | GemmMicrokernelTester() |
| 10569 | .mr(4) |
| 10570 | .nr(16) |
| 10571 | .kr(1) |
| 10572 | .sr(1) |
| 10573 | .m(4) |
| 10574 | .n(16) |
| 10575 | .k(8) |
| 10576 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10577 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10578 | } |
| 10579 | |
| 10580 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, no_a_zero_point) { |
| 10581 | TEST_REQUIRES_ARM_NEON; |
| 10582 | for (size_t k = 1; k <= 40; k += 9) { |
| 10583 | GemmMicrokernelTester() |
| 10584 | .mr(4) |
| 10585 | .nr(16) |
| 10586 | .kr(1) |
| 10587 | .sr(1) |
| 10588 | .m(4) |
| 10589 | .n(16) |
| 10590 | .k(k) |
| 10591 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10592 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10593 | } |
| 10594 | } |
| 10595 | |
| 10596 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, no_b_zero_point) { |
| 10597 | TEST_REQUIRES_ARM_NEON; |
| 10598 | for (size_t k = 1; k <= 40; k += 9) { |
| 10599 | GemmMicrokernelTester() |
| 10600 | .mr(4) |
| 10601 | .nr(16) |
| 10602 | .kr(1) |
| 10603 | .sr(1) |
| 10604 | .m(4) |
| 10605 | .n(16) |
| 10606 | .k(k) |
| 10607 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10608 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10609 | } |
| 10610 | } |
| 10611 | |
| 10612 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A75, no_zero_point) { |
| 10613 | TEST_REQUIRES_ARM_NEON; |
| 10614 | for (size_t k = 1; k <= 40; k += 9) { |
| 10615 | GemmMicrokernelTester() |
| 10616 | .mr(4) |
| 10617 | .nr(16) |
| 10618 | .kr(1) |
| 10619 | .sr(1) |
| 10620 | .m(4) |
| 10621 | .n(16) |
| 10622 | .k(k) |
| 10623 | .a_zero_point(0) |
| 10624 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10625 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a75, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10626 | } |
| 10627 | } |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 10628 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | fb3a94f | 2021-08-02 20:37:06 -0700 | [diff] [blame] | 10629 | |
| 10630 | |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 10631 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10632 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) { |
| 10633 | TEST_REQUIRES_ARM_NEON; |
| 10634 | GemmMicrokernelTester() |
| 10635 | .mr(4) |
| 10636 | .nr(16) |
| 10637 | .kr(1) |
| 10638 | .sr(1) |
| 10639 | .m(4) |
| 10640 | .n(16) |
| 10641 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10642 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10643 | } |
| 10644 | |
| 10645 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, strided_cn) { |
| 10646 | TEST_REQUIRES_ARM_NEON; |
| 10647 | GemmMicrokernelTester() |
| 10648 | .mr(4) |
| 10649 | .nr(16) |
| 10650 | .kr(1) |
| 10651 | .sr(1) |
| 10652 | .m(4) |
| 10653 | .n(16) |
| 10654 | .k(8) |
| 10655 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10656 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10657 | } |
| 10658 | |
| 10659 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_strided_a) { |
| 10660 | TEST_REQUIRES_ARM_NEON; |
| 10661 | GemmMicrokernelTester() |
| 10662 | .mr(4) |
| 10663 | .nr(16) |
| 10664 | .kr(1) |
| 10665 | .sr(1) |
| 10666 | .m(4) |
| 10667 | .n(16) |
| 10668 | .k(8) |
| 10669 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10670 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10671 | } |
| 10672 | |
| 10673 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_subtile) { |
| 10674 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10675 | for (uint32_t n = 1; n <= 16; n++) { |
| 10676 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10677 | GemmMicrokernelTester() |
| 10678 | .mr(4) |
| 10679 | .nr(16) |
| 10680 | .kr(1) |
| 10681 | .sr(1) |
| 10682 | .m(m) |
| 10683 | .n(n) |
| 10684 | .k(8) |
| 10685 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10686 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10687 | } |
| 10688 | } |
| 10689 | } |
| 10690 | |
| 10691 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_subtile_m) { |
| 10692 | TEST_REQUIRES_ARM_NEON; |
| 10693 | for (uint32_t m = 1; m <= 4; m++) { |
| 10694 | GemmMicrokernelTester() |
| 10695 | .mr(4) |
| 10696 | .nr(16) |
| 10697 | .kr(1) |
| 10698 | .sr(1) |
| 10699 | .m(m) |
| 10700 | .n(16) |
| 10701 | .k(8) |
| 10702 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10703 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10704 | } |
| 10705 | } |
| 10706 | |
| 10707 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8_subtile_n) { |
| 10708 | TEST_REQUIRES_ARM_NEON; |
| 10709 | for (uint32_t n = 1; n <= 16; n++) { |
| 10710 | GemmMicrokernelTester() |
| 10711 | .mr(4) |
| 10712 | .nr(16) |
| 10713 | .kr(1) |
| 10714 | .sr(1) |
| 10715 | .m(4) |
| 10716 | .n(n) |
| 10717 | .k(8) |
| 10718 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10719 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10720 | } |
| 10721 | } |
| 10722 | |
| 10723 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_lt_8) { |
| 10724 | TEST_REQUIRES_ARM_NEON; |
| 10725 | for (size_t k = 1; k < 8; k++) { |
| 10726 | GemmMicrokernelTester() |
| 10727 | .mr(4) |
| 10728 | .nr(16) |
| 10729 | .kr(1) |
| 10730 | .sr(1) |
| 10731 | .m(4) |
| 10732 | .n(16) |
| 10733 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10734 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10735 | } |
| 10736 | } |
| 10737 | |
| 10738 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_lt_8_strided_a) { |
| 10739 | TEST_REQUIRES_ARM_NEON; |
| 10740 | for (size_t k = 1; k < 8; k++) { |
| 10741 | GemmMicrokernelTester() |
| 10742 | .mr(4) |
| 10743 | .nr(16) |
| 10744 | .kr(1) |
| 10745 | .sr(1) |
| 10746 | .m(4) |
| 10747 | .n(16) |
| 10748 | .k(k) |
| 10749 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10750 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10751 | } |
| 10752 | } |
| 10753 | |
| 10754 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_lt_8_subtile) { |
| 10755 | TEST_REQUIRES_ARM_NEON; |
| 10756 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10757 | for (uint32_t n = 1; n <= 16; n++) { |
| 10758 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10759 | GemmMicrokernelTester() |
| 10760 | .mr(4) |
| 10761 | .nr(16) |
| 10762 | .kr(1) |
| 10763 | .sr(1) |
| 10764 | .m(m) |
| 10765 | .n(n) |
| 10766 | .k(k) |
| 10767 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10768 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10769 | } |
| 10770 | } |
| 10771 | } |
| 10772 | } |
| 10773 | |
| 10774 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_gt_8) { |
| 10775 | TEST_REQUIRES_ARM_NEON; |
| 10776 | for (size_t k = 9; k < 16; k++) { |
| 10777 | GemmMicrokernelTester() |
| 10778 | .mr(4) |
| 10779 | .nr(16) |
| 10780 | .kr(1) |
| 10781 | .sr(1) |
| 10782 | .m(4) |
| 10783 | .n(16) |
| 10784 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10785 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10786 | } |
| 10787 | } |
| 10788 | |
| 10789 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_gt_8_strided_a) { |
| 10790 | TEST_REQUIRES_ARM_NEON; |
| 10791 | for (size_t k = 9; k < 16; k++) { |
| 10792 | GemmMicrokernelTester() |
| 10793 | .mr(4) |
| 10794 | .nr(16) |
| 10795 | .kr(1) |
| 10796 | .sr(1) |
| 10797 | .m(4) |
| 10798 | .n(16) |
| 10799 | .k(k) |
| 10800 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10801 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10802 | } |
| 10803 | } |
| 10804 | |
| 10805 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_gt_8_subtile) { |
| 10806 | TEST_REQUIRES_ARM_NEON; |
| 10807 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10808 | for (uint32_t n = 1; n <= 16; n++) { |
| 10809 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10810 | GemmMicrokernelTester() |
| 10811 | .mr(4) |
| 10812 | .nr(16) |
| 10813 | .kr(1) |
| 10814 | .sr(1) |
| 10815 | .m(m) |
| 10816 | .n(n) |
| 10817 | .k(k) |
| 10818 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10819 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10820 | } |
| 10821 | } |
| 10822 | } |
| 10823 | } |
| 10824 | |
| 10825 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_div_8) { |
| 10826 | TEST_REQUIRES_ARM_NEON; |
| 10827 | for (size_t k = 16; k <= 80; k += 8) { |
| 10828 | GemmMicrokernelTester() |
| 10829 | .mr(4) |
| 10830 | .nr(16) |
| 10831 | .kr(1) |
| 10832 | .sr(1) |
| 10833 | .m(4) |
| 10834 | .n(16) |
| 10835 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10836 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10837 | } |
| 10838 | } |
| 10839 | |
| 10840 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_div_8_strided_a) { |
| 10841 | TEST_REQUIRES_ARM_NEON; |
| 10842 | for (size_t k = 16; k <= 80; k += 8) { |
| 10843 | GemmMicrokernelTester() |
| 10844 | .mr(4) |
| 10845 | .nr(16) |
| 10846 | .kr(1) |
| 10847 | .sr(1) |
| 10848 | .m(4) |
| 10849 | .n(16) |
| 10850 | .k(k) |
| 10851 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10852 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10853 | } |
| 10854 | } |
| 10855 | |
| 10856 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_div_8_subtile) { |
| 10857 | TEST_REQUIRES_ARM_NEON; |
| 10858 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10859 | for (uint32_t n = 1; n <= 16; n++) { |
| 10860 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10861 | GemmMicrokernelTester() |
| 10862 | .mr(4) |
| 10863 | .nr(16) |
| 10864 | .kr(1) |
| 10865 | .sr(1) |
| 10866 | .m(m) |
| 10867 | .n(n) |
| 10868 | .k(k) |
| 10869 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10870 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10871 | } |
| 10872 | } |
| 10873 | } |
| 10874 | } |
| 10875 | |
| 10876 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_16) { |
| 10877 | TEST_REQUIRES_ARM_NEON; |
| 10878 | for (uint32_t n = 17; n < 32; n++) { |
| 10879 | for (size_t k = 1; k <= 40; k += 9) { |
| 10880 | GemmMicrokernelTester() |
| 10881 | .mr(4) |
| 10882 | .nr(16) |
| 10883 | .kr(1) |
| 10884 | .sr(1) |
| 10885 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10886 | .n(n) |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10887 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10888 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10889 | } |
| 10890 | } |
| 10891 | } |
| 10892 | |
| 10893 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_16_strided_cn) { |
| 10894 | TEST_REQUIRES_ARM_NEON; |
| 10895 | for (uint32_t n = 17; n < 32; n++) { |
| 10896 | for (size_t k = 1; k <= 40; k += 9) { |
| 10897 | GemmMicrokernelTester() |
| 10898 | .mr(4) |
| 10899 | .nr(16) |
| 10900 | .kr(1) |
| 10901 | .sr(1) |
| 10902 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10903 | .n(n) |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10904 | .k(k) |
| 10905 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10906 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10907 | } |
| 10908 | } |
| 10909 | } |
| 10910 | |
| 10911 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_16_strided_a) { |
| 10912 | TEST_REQUIRES_ARM_NEON; |
| 10913 | for (uint32_t n = 17; n < 32; n++) { |
| 10914 | for (size_t k = 1; k <= 40; k += 9) { |
| 10915 | GemmMicrokernelTester() |
| 10916 | .mr(4) |
| 10917 | .nr(16) |
| 10918 | .kr(1) |
| 10919 | .sr(1) |
| 10920 | .m(4) |
| 10921 | .n(n) |
| 10922 | .k(k) |
| 10923 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10924 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10925 | } |
| 10926 | } |
| 10927 | } |
| 10928 | |
| 10929 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_gt_16_subtile) { |
| 10930 | TEST_REQUIRES_ARM_NEON; |
| 10931 | for (uint32_t n = 17; n < 32; n++) { |
| 10932 | for (size_t k = 1; k <= 40; k += 9) { |
| 10933 | for (uint32_t m = 1; m <= 4; m++) { |
| 10934 | GemmMicrokernelTester() |
| 10935 | .mr(4) |
| 10936 | .nr(16) |
| 10937 | .kr(1) |
| 10938 | .sr(1) |
| 10939 | .m(m) |
| 10940 | .n(n) |
| 10941 | .k(k) |
| 10942 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10943 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10944 | } |
| 10945 | } |
| 10946 | } |
| 10947 | } |
| 10948 | |
| 10949 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_16) { |
| 10950 | TEST_REQUIRES_ARM_NEON; |
| 10951 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10952 | for (size_t k = 1; k <= 40; k += 9) { |
| 10953 | GemmMicrokernelTester() |
| 10954 | .mr(4) |
| 10955 | .nr(16) |
| 10956 | .kr(1) |
| 10957 | .sr(1) |
| 10958 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10959 | .n(n) |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10960 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10961 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10962 | } |
| 10963 | } |
| 10964 | } |
| 10965 | |
| 10966 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_16_strided_cn) { |
| 10967 | TEST_REQUIRES_ARM_NEON; |
| 10968 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10969 | for (size_t k = 1; k <= 40; k += 9) { |
| 10970 | GemmMicrokernelTester() |
| 10971 | .mr(4) |
| 10972 | .nr(16) |
| 10973 | .kr(1) |
| 10974 | .sr(1) |
| 10975 | .m(4) |
| 10976 | .n(n) |
| 10977 | .k(k) |
| 10978 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10979 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10980 | } |
| 10981 | } |
| 10982 | } |
| 10983 | |
| 10984 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_16_strided_a) { |
| 10985 | TEST_REQUIRES_ARM_NEON; |
| 10986 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10987 | for (size_t k = 1; k <= 40; k += 9) { |
| 10988 | GemmMicrokernelTester() |
| 10989 | .mr(4) |
| 10990 | .nr(16) |
| 10991 | .kr(1) |
| 10992 | .sr(1) |
| 10993 | .m(4) |
| 10994 | .n(n) |
| 10995 | .k(k) |
| 10996 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10997 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 10998 | } |
| 10999 | } |
| 11000 | } |
| 11001 | |
| 11002 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, n_div_16_subtile) { |
| 11003 | TEST_REQUIRES_ARM_NEON; |
| 11004 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 11005 | for (size_t k = 1; k <= 40; k += 9) { |
| 11006 | for (uint32_t m = 1; m <= 4; m++) { |
| 11007 | GemmMicrokernelTester() |
| 11008 | .mr(4) |
| 11009 | .nr(16) |
| 11010 | .kr(1) |
| 11011 | .sr(1) |
| 11012 | .m(m) |
| 11013 | .n(n) |
| 11014 | .k(k) |
| 11015 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11016 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11017 | } |
| 11018 | } |
| 11019 | } |
| 11020 | } |
| 11021 | |
| 11022 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, strided_cm_subtile) { |
| 11023 | TEST_REQUIRES_ARM_NEON; |
| 11024 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11025 | for (uint32_t n = 1; n <= 16; n++) { |
| 11026 | for (uint32_t m = 1; m <= 4; m++) { |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11027 | GemmMicrokernelTester() |
| 11028 | .mr(4) |
| 11029 | .nr(16) |
| 11030 | .kr(1) |
| 11031 | .sr(1) |
| 11032 | .m(m) |
| 11033 | .n(n) |
| 11034 | .k(k) |
| 11035 | .cm_stride(19) |
| 11036 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11037 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11038 | } |
| 11039 | } |
| 11040 | } |
| 11041 | } |
| 11042 | |
| 11043 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, qmin) { |
| 11044 | TEST_REQUIRES_ARM_NEON; |
| 11045 | GemmMicrokernelTester() |
| 11046 | .mr(4) |
| 11047 | .nr(16) |
| 11048 | .kr(1) |
| 11049 | .sr(1) |
| 11050 | .m(4) |
| 11051 | .n(16) |
| 11052 | .k(8) |
| 11053 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11054 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11055 | } |
| 11056 | |
| 11057 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, qmax) { |
| 11058 | TEST_REQUIRES_ARM_NEON; |
| 11059 | GemmMicrokernelTester() |
| 11060 | .mr(4) |
| 11061 | .nr(16) |
| 11062 | .kr(1) |
| 11063 | .sr(1) |
| 11064 | .m(4) |
| 11065 | .n(16) |
| 11066 | .k(8) |
| 11067 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11068 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11069 | } |
| 11070 | |
| 11071 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, strided_cm) { |
| 11072 | TEST_REQUIRES_ARM_NEON; |
| 11073 | GemmMicrokernelTester() |
| 11074 | .mr(4) |
| 11075 | .nr(16) |
| 11076 | .kr(1) |
| 11077 | .sr(1) |
| 11078 | .m(4) |
| 11079 | .n(16) |
| 11080 | .k(8) |
| 11081 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11082 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11083 | } |
| 11084 | |
| 11085 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, no_a_zero_point) { |
| 11086 | TEST_REQUIRES_ARM_NEON; |
| 11087 | for (size_t k = 1; k <= 40; k += 9) { |
| 11088 | GemmMicrokernelTester() |
| 11089 | .mr(4) |
| 11090 | .nr(16) |
| 11091 | .kr(1) |
| 11092 | .sr(1) |
| 11093 | .m(4) |
| 11094 | .n(16) |
| 11095 | .k(k) |
| 11096 | .a_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11097 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11098 | } |
| 11099 | } |
| 11100 | |
| 11101 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, no_b_zero_point) { |
| 11102 | TEST_REQUIRES_ARM_NEON; |
| 11103 | for (size_t k = 1; k <= 40; k += 9) { |
| 11104 | GemmMicrokernelTester() |
| 11105 | .mr(4) |
| 11106 | .nr(16) |
| 11107 | .kr(1) |
| 11108 | .sr(1) |
| 11109 | .m(4) |
| 11110 | .n(16) |
| 11111 | .k(k) |
| 11112 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11113 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11114 | } |
| 11115 | } |
| 11116 | |
| 11117 | TEST(QU8_GEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, no_zero_point) { |
| 11118 | TEST_REQUIRES_ARM_NEON; |
| 11119 | for (size_t k = 1; k <= 40; k += 9) { |
| 11120 | GemmMicrokernelTester() |
| 11121 | .mr(4) |
| 11122 | .nr(16) |
| 11123 | .kr(1) |
| 11124 | .sr(1) |
| 11125 | .m(4) |
| 11126 | .n(16) |
| 11127 | .k(k) |
| 11128 | .a_zero_point(0) |
| 11129 | .b_zero_point(0) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11130 | .Test(xnn_qu8_gemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53, xnn_init_qu8_conv_minmax_rndnu_neon_params, xnn_qu8_requantize_rndnu); |
Frank Barchard | 59ed1da | 2021-08-02 11:34:59 -0700 | [diff] [blame] | 11131 | } |
| 11132 | } |
Frank Barchard | e4d3f76 | 2021-12-23 15:31:43 -0800 | [diff] [blame] | 11133 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |