Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [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/qs8-gemm-minmax-rndnu.yaml |
| 11 | // Generator: tools/generate-gemm-test.py |
| 12 | |
| 13 | |
| 14 | #include <gtest/gtest.h> |
| 15 | |
| 16 | #include <xnnpack/allocator.h> |
| 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(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 38 | } |
| 39 | |
| 40 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 52 | } |
| 53 | |
| 54 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 66 | } |
| 67 | |
| 68 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 82 | } |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 99 | } |
| 100 | } |
| 101 | |
| 102 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 115 | } |
| 116 | } |
| 117 | |
| 118 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 181 | } |
| 182 | } |
| 183 | |
| 184 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 197 | } |
| 198 | } |
| 199 | |
| 200 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 232 | } |
| 233 | } |
| 234 | |
| 235 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 266 | } |
| 267 | } |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 284 | } |
| 285 | } |
| 286 | } |
| 287 | |
| 288 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 302 | } |
| 303 | } |
| 304 | } |
| 305 | |
| 306 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 320 | } |
| 321 | } |
| 322 | } |
| 323 | |
| 324 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 357 | } |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 375 | } |
| 376 | } |
| 377 | } |
| 378 | |
| 379 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 393 | } |
| 394 | } |
| 395 | } |
| 396 | |
| 397 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 412 | } |
| 413 | } |
| 414 | } |
| 415 | } |
| 416 | |
| 417 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 433 | } |
| 434 | } |
| 435 | } |
| 436 | } |
| 437 | |
| 438 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 450 | } |
| 451 | |
| 452 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 464 | } |
| 465 | |
| 466 | TEST(QS8_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_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
| 478 | } |
| 479 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY |
| 480 | |
| 481 | |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 482 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 483 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8) { |
| 484 | TEST_REQUIRES_ARM_NEON; |
| 485 | GemmMicrokernelTester() |
| 486 | .mr(4) |
| 487 | .nr(8) |
| 488 | .kr(1) |
| 489 | .sr(1) |
| 490 | .m(4) |
| 491 | .n(8) |
| 492 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 493 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 494 | } |
| 495 | |
| 496 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, strided_cn) { |
| 497 | TEST_REQUIRES_ARM_NEON; |
| 498 | GemmMicrokernelTester() |
| 499 | .mr(4) |
| 500 | .nr(8) |
| 501 | .kr(1) |
| 502 | .sr(1) |
| 503 | .m(4) |
| 504 | .n(8) |
| 505 | .k(8) |
| 506 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 507 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 508 | } |
| 509 | |
| 510 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8_strided_a) { |
| 511 | TEST_REQUIRES_ARM_NEON; |
| 512 | GemmMicrokernelTester() |
| 513 | .mr(4) |
| 514 | .nr(8) |
| 515 | .kr(1) |
| 516 | .sr(1) |
| 517 | .m(4) |
| 518 | .n(8) |
| 519 | .k(8) |
| 520 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 521 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 522 | } |
| 523 | |
| 524 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8_subtile) { |
| 525 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 526 | for (uint32_t n = 1; n <= 8; n++) { |
| 527 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 528 | GemmMicrokernelTester() |
| 529 | .mr(4) |
| 530 | .nr(8) |
| 531 | .kr(1) |
| 532 | .sr(1) |
| 533 | .m(m) |
| 534 | .n(n) |
| 535 | .k(8) |
| 536 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 537 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 538 | } |
| 539 | } |
| 540 | } |
| 541 | |
| 542 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8_subtile_m) { |
| 543 | TEST_REQUIRES_ARM_NEON; |
| 544 | for (uint32_t m = 1; m <= 4; m++) { |
| 545 | GemmMicrokernelTester() |
| 546 | .mr(4) |
| 547 | .nr(8) |
| 548 | .kr(1) |
| 549 | .sr(1) |
| 550 | .m(m) |
| 551 | .n(8) |
| 552 | .k(8) |
| 553 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 554 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 555 | } |
| 556 | } |
| 557 | |
| 558 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8_subtile_n) { |
| 559 | TEST_REQUIRES_ARM_NEON; |
| 560 | for (uint32_t n = 1; n <= 8; n++) { |
| 561 | GemmMicrokernelTester() |
| 562 | .mr(4) |
| 563 | .nr(8) |
| 564 | .kr(1) |
| 565 | .sr(1) |
| 566 | .m(4) |
| 567 | .n(n) |
| 568 | .k(8) |
| 569 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 570 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 571 | } |
| 572 | } |
| 573 | |
| 574 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_lt_8) { |
| 575 | TEST_REQUIRES_ARM_NEON; |
| 576 | for (size_t k = 1; k < 8; k++) { |
| 577 | GemmMicrokernelTester() |
| 578 | .mr(4) |
| 579 | .nr(8) |
| 580 | .kr(1) |
| 581 | .sr(1) |
| 582 | .m(4) |
| 583 | .n(8) |
| 584 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 585 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 586 | } |
| 587 | } |
| 588 | |
| 589 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_lt_8_strided_a) { |
| 590 | TEST_REQUIRES_ARM_NEON; |
| 591 | for (size_t k = 1; k < 8; k++) { |
| 592 | GemmMicrokernelTester() |
| 593 | .mr(4) |
| 594 | .nr(8) |
| 595 | .kr(1) |
| 596 | .sr(1) |
| 597 | .m(4) |
| 598 | .n(8) |
| 599 | .k(k) |
| 600 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 601 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 602 | } |
| 603 | } |
| 604 | |
| 605 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_lt_8_subtile) { |
| 606 | TEST_REQUIRES_ARM_NEON; |
| 607 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 608 | for (uint32_t n = 1; n <= 8; n++) { |
| 609 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 610 | GemmMicrokernelTester() |
| 611 | .mr(4) |
| 612 | .nr(8) |
| 613 | .kr(1) |
| 614 | .sr(1) |
| 615 | .m(m) |
| 616 | .n(n) |
| 617 | .k(k) |
| 618 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 619 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 620 | } |
| 621 | } |
| 622 | } |
| 623 | } |
| 624 | |
| 625 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_gt_8) { |
| 626 | TEST_REQUIRES_ARM_NEON; |
| 627 | for (size_t k = 9; k < 16; k++) { |
| 628 | GemmMicrokernelTester() |
| 629 | .mr(4) |
| 630 | .nr(8) |
| 631 | .kr(1) |
| 632 | .sr(1) |
| 633 | .m(4) |
| 634 | .n(8) |
| 635 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 636 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 637 | } |
| 638 | } |
| 639 | |
| 640 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_gt_8_strided_a) { |
| 641 | TEST_REQUIRES_ARM_NEON; |
| 642 | for (size_t k = 9; k < 16; k++) { |
| 643 | GemmMicrokernelTester() |
| 644 | .mr(4) |
| 645 | .nr(8) |
| 646 | .kr(1) |
| 647 | .sr(1) |
| 648 | .m(4) |
| 649 | .n(8) |
| 650 | .k(k) |
| 651 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 652 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 653 | } |
| 654 | } |
| 655 | |
| 656 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_gt_8_subtile) { |
| 657 | TEST_REQUIRES_ARM_NEON; |
| 658 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 659 | for (uint32_t n = 1; n <= 8; n++) { |
| 660 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 661 | GemmMicrokernelTester() |
| 662 | .mr(4) |
| 663 | .nr(8) |
| 664 | .kr(1) |
| 665 | .sr(1) |
| 666 | .m(m) |
| 667 | .n(n) |
| 668 | .k(k) |
| 669 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 670 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 671 | } |
| 672 | } |
| 673 | } |
| 674 | } |
| 675 | |
| 676 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_div_8) { |
| 677 | TEST_REQUIRES_ARM_NEON; |
| 678 | for (size_t k = 16; k <= 80; k += 8) { |
| 679 | GemmMicrokernelTester() |
| 680 | .mr(4) |
| 681 | .nr(8) |
| 682 | .kr(1) |
| 683 | .sr(1) |
| 684 | .m(4) |
| 685 | .n(8) |
| 686 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 687 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 688 | } |
| 689 | } |
| 690 | |
| 691 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_div_8_strided_a) { |
| 692 | TEST_REQUIRES_ARM_NEON; |
| 693 | for (size_t k = 16; k <= 80; k += 8) { |
| 694 | GemmMicrokernelTester() |
| 695 | .mr(4) |
| 696 | .nr(8) |
| 697 | .kr(1) |
| 698 | .sr(1) |
| 699 | .m(4) |
| 700 | .n(8) |
| 701 | .k(k) |
| 702 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 703 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 704 | } |
| 705 | } |
| 706 | |
| 707 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_div_8_subtile) { |
| 708 | TEST_REQUIRES_ARM_NEON; |
| 709 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 710 | for (uint32_t n = 1; n <= 8; n++) { |
| 711 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 712 | GemmMicrokernelTester() |
| 713 | .mr(4) |
| 714 | .nr(8) |
| 715 | .kr(1) |
| 716 | .sr(1) |
| 717 | .m(m) |
| 718 | .n(n) |
| 719 | .k(k) |
| 720 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 721 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 722 | } |
| 723 | } |
| 724 | } |
| 725 | } |
| 726 | |
| 727 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8) { |
| 728 | TEST_REQUIRES_ARM_NEON; |
| 729 | for (uint32_t n = 9; n < 16; n++) { |
| 730 | for (size_t k = 1; k <= 40; k += 9) { |
| 731 | GemmMicrokernelTester() |
| 732 | .mr(4) |
| 733 | .nr(8) |
| 734 | .kr(1) |
| 735 | .sr(1) |
| 736 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 737 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 738 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 739 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 740 | } |
| 741 | } |
| 742 | } |
| 743 | |
| 744 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8_strided_cn) { |
| 745 | TEST_REQUIRES_ARM_NEON; |
| 746 | for (uint32_t n = 9; n < 16; n++) { |
| 747 | for (size_t k = 1; k <= 40; k += 9) { |
| 748 | GemmMicrokernelTester() |
| 749 | .mr(4) |
| 750 | .nr(8) |
| 751 | .kr(1) |
| 752 | .sr(1) |
| 753 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 754 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 755 | .k(k) |
| 756 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 757 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 758 | } |
| 759 | } |
| 760 | } |
| 761 | |
| 762 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8_strided_a) { |
| 763 | TEST_REQUIRES_ARM_NEON; |
| 764 | for (uint32_t n = 9; n < 16; n++) { |
| 765 | for (size_t k = 1; k <= 40; k += 9) { |
| 766 | GemmMicrokernelTester() |
| 767 | .mr(4) |
| 768 | .nr(8) |
| 769 | .kr(1) |
| 770 | .sr(1) |
| 771 | .m(4) |
| 772 | .n(n) |
| 773 | .k(k) |
| 774 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 775 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 776 | } |
| 777 | } |
| 778 | } |
| 779 | |
| 780 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8_subtile) { |
| 781 | TEST_REQUIRES_ARM_NEON; |
| 782 | for (uint32_t n = 9; n < 16; n++) { |
| 783 | for (size_t k = 1; k <= 40; k += 9) { |
| 784 | for (uint32_t m = 1; m <= 4; m++) { |
| 785 | GemmMicrokernelTester() |
| 786 | .mr(4) |
| 787 | .nr(8) |
| 788 | .kr(1) |
| 789 | .sr(1) |
| 790 | .m(m) |
| 791 | .n(n) |
| 792 | .k(k) |
| 793 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 794 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 795 | } |
| 796 | } |
| 797 | } |
| 798 | } |
| 799 | |
| 800 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8) { |
| 801 | TEST_REQUIRES_ARM_NEON; |
| 802 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 803 | for (size_t k = 1; k <= 40; k += 9) { |
| 804 | GemmMicrokernelTester() |
| 805 | .mr(4) |
| 806 | .nr(8) |
| 807 | .kr(1) |
| 808 | .sr(1) |
| 809 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 810 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 811 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 812 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 813 | } |
| 814 | } |
| 815 | } |
| 816 | |
| 817 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8_strided_cn) { |
| 818 | TEST_REQUIRES_ARM_NEON; |
| 819 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 820 | for (size_t k = 1; k <= 40; k += 9) { |
| 821 | GemmMicrokernelTester() |
| 822 | .mr(4) |
| 823 | .nr(8) |
| 824 | .kr(1) |
| 825 | .sr(1) |
| 826 | .m(4) |
| 827 | .n(n) |
| 828 | .k(k) |
| 829 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 830 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 831 | } |
| 832 | } |
| 833 | } |
| 834 | |
| 835 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8_strided_a) { |
| 836 | TEST_REQUIRES_ARM_NEON; |
| 837 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 838 | for (size_t k = 1; k <= 40; k += 9) { |
| 839 | GemmMicrokernelTester() |
| 840 | .mr(4) |
| 841 | .nr(8) |
| 842 | .kr(1) |
| 843 | .sr(1) |
| 844 | .m(4) |
| 845 | .n(n) |
| 846 | .k(k) |
| 847 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 848 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 849 | } |
| 850 | } |
| 851 | } |
| 852 | |
| 853 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8_subtile) { |
| 854 | TEST_REQUIRES_ARM_NEON; |
| 855 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 856 | for (size_t k = 1; k <= 40; k += 9) { |
| 857 | for (uint32_t m = 1; m <= 4; m++) { |
| 858 | GemmMicrokernelTester() |
| 859 | .mr(4) |
| 860 | .nr(8) |
| 861 | .kr(1) |
| 862 | .sr(1) |
| 863 | .m(m) |
| 864 | .n(n) |
| 865 | .k(k) |
| 866 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 867 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 868 | } |
| 869 | } |
| 870 | } |
| 871 | } |
| 872 | |
| 873 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, strided_cm_subtile) { |
| 874 | TEST_REQUIRES_ARM_NEON; |
| 875 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 876 | for (uint32_t n = 1; n <= 8; n++) { |
| 877 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 878 | GemmMicrokernelTester() |
| 879 | .mr(4) |
| 880 | .nr(8) |
| 881 | .kr(1) |
| 882 | .sr(1) |
| 883 | .m(m) |
| 884 | .n(n) |
| 885 | .k(k) |
| 886 | .cm_stride(11) |
| 887 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 888 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 889 | } |
| 890 | } |
| 891 | } |
| 892 | } |
| 893 | |
| 894 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, qmin) { |
| 895 | TEST_REQUIRES_ARM_NEON; |
| 896 | GemmMicrokernelTester() |
| 897 | .mr(4) |
| 898 | .nr(8) |
| 899 | .kr(1) |
| 900 | .sr(1) |
| 901 | .m(4) |
| 902 | .n(8) |
| 903 | .k(8) |
| 904 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 905 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 906 | } |
| 907 | |
| 908 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, qmax) { |
| 909 | TEST_REQUIRES_ARM_NEON; |
| 910 | GemmMicrokernelTester() |
| 911 | .mr(4) |
| 912 | .nr(8) |
| 913 | .kr(1) |
| 914 | .sr(1) |
| 915 | .m(4) |
| 916 | .n(8) |
| 917 | .k(8) |
| 918 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 919 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 920 | } |
| 921 | |
| 922 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, strided_cm) { |
| 923 | TEST_REQUIRES_ARM_NEON; |
| 924 | GemmMicrokernelTester() |
| 925 | .mr(4) |
| 926 | .nr(8) |
| 927 | .kr(1) |
| 928 | .sr(1) |
| 929 | .m(4) |
| 930 | .n(8) |
| 931 | .k(8) |
| 932 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 933 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 934 | } |
| 935 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 936 | |
| 937 | |
| 938 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 939 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8) { |
| 940 | TEST_REQUIRES_ARM_NEON; |
| 941 | GemmMicrokernelTester() |
| 942 | .mr(1) |
| 943 | .nr(8) |
| 944 | .kr(2) |
| 945 | .sr(1) |
| 946 | .m(1) |
| 947 | .n(8) |
| 948 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 949 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 950 | } |
| 951 | |
| 952 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, strided_cn) { |
| 953 | TEST_REQUIRES_ARM_NEON; |
| 954 | GemmMicrokernelTester() |
| 955 | .mr(1) |
| 956 | .nr(8) |
| 957 | .kr(2) |
| 958 | .sr(1) |
| 959 | .m(1) |
| 960 | .n(8) |
| 961 | .k(8) |
| 962 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 963 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 964 | } |
| 965 | |
| 966 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_strided_a) { |
| 967 | TEST_REQUIRES_ARM_NEON; |
| 968 | GemmMicrokernelTester() |
| 969 | .mr(1) |
| 970 | .nr(8) |
| 971 | .kr(2) |
| 972 | .sr(1) |
| 973 | .m(1) |
| 974 | .n(8) |
| 975 | .k(8) |
| 976 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 977 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 978 | } |
| 979 | |
| 980 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile) { |
| 981 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 982 | for (uint32_t n = 1; n <= 8; n++) { |
| 983 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 984 | GemmMicrokernelTester() |
| 985 | .mr(1) |
| 986 | .nr(8) |
| 987 | .kr(2) |
| 988 | .sr(1) |
| 989 | .m(m) |
| 990 | .n(n) |
| 991 | .k(8) |
| 992 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 993 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 994 | } |
| 995 | } |
| 996 | } |
| 997 | |
| 998 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile_m) { |
| 999 | TEST_REQUIRES_ARM_NEON; |
| 1000 | for (uint32_t m = 1; m <= 1; m++) { |
| 1001 | GemmMicrokernelTester() |
| 1002 | .mr(1) |
| 1003 | .nr(8) |
| 1004 | .kr(2) |
| 1005 | .sr(1) |
| 1006 | .m(m) |
| 1007 | .n(8) |
| 1008 | .k(8) |
| 1009 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1010 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1011 | } |
| 1012 | } |
| 1013 | |
| 1014 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile_n) { |
| 1015 | TEST_REQUIRES_ARM_NEON; |
| 1016 | for (uint32_t n = 1; n <= 8; n++) { |
| 1017 | GemmMicrokernelTester() |
| 1018 | .mr(1) |
| 1019 | .nr(8) |
| 1020 | .kr(2) |
| 1021 | .sr(1) |
| 1022 | .m(1) |
| 1023 | .n(n) |
| 1024 | .k(8) |
| 1025 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1026 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1027 | } |
| 1028 | } |
| 1029 | |
| 1030 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_lt_8) { |
| 1031 | TEST_REQUIRES_ARM_NEON; |
| 1032 | for (size_t k = 1; k < 8; k++) { |
| 1033 | GemmMicrokernelTester() |
| 1034 | .mr(1) |
| 1035 | .nr(8) |
| 1036 | .kr(2) |
| 1037 | .sr(1) |
| 1038 | .m(1) |
| 1039 | .n(8) |
| 1040 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1041 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1042 | } |
| 1043 | } |
| 1044 | |
| 1045 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_lt_8_strided_a) { |
| 1046 | TEST_REQUIRES_ARM_NEON; |
| 1047 | for (size_t k = 1; k < 8; k++) { |
| 1048 | GemmMicrokernelTester() |
| 1049 | .mr(1) |
| 1050 | .nr(8) |
| 1051 | .kr(2) |
| 1052 | .sr(1) |
| 1053 | .m(1) |
| 1054 | .n(8) |
| 1055 | .k(k) |
| 1056 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1057 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1058 | } |
| 1059 | } |
| 1060 | |
| 1061 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_lt_8_subtile) { |
| 1062 | TEST_REQUIRES_ARM_NEON; |
| 1063 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1064 | for (uint32_t n = 1; n <= 8; n++) { |
| 1065 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1066 | GemmMicrokernelTester() |
| 1067 | .mr(1) |
| 1068 | .nr(8) |
| 1069 | .kr(2) |
| 1070 | .sr(1) |
| 1071 | .m(m) |
| 1072 | .n(n) |
| 1073 | .k(k) |
| 1074 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1075 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1076 | } |
| 1077 | } |
| 1078 | } |
| 1079 | } |
| 1080 | |
| 1081 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_gt_8) { |
| 1082 | TEST_REQUIRES_ARM_NEON; |
| 1083 | for (size_t k = 9; k < 16; k++) { |
| 1084 | GemmMicrokernelTester() |
| 1085 | .mr(1) |
| 1086 | .nr(8) |
| 1087 | .kr(2) |
| 1088 | .sr(1) |
| 1089 | .m(1) |
| 1090 | .n(8) |
| 1091 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1092 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1093 | } |
| 1094 | } |
| 1095 | |
| 1096 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_gt_8_strided_a) { |
| 1097 | TEST_REQUIRES_ARM_NEON; |
| 1098 | for (size_t k = 9; k < 16; k++) { |
| 1099 | GemmMicrokernelTester() |
| 1100 | .mr(1) |
| 1101 | .nr(8) |
| 1102 | .kr(2) |
| 1103 | .sr(1) |
| 1104 | .m(1) |
| 1105 | .n(8) |
| 1106 | .k(k) |
| 1107 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1108 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1109 | } |
| 1110 | } |
| 1111 | |
| 1112 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_gt_8_subtile) { |
| 1113 | TEST_REQUIRES_ARM_NEON; |
| 1114 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1115 | for (uint32_t n = 1; n <= 8; n++) { |
| 1116 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1117 | GemmMicrokernelTester() |
| 1118 | .mr(1) |
| 1119 | .nr(8) |
| 1120 | .kr(2) |
| 1121 | .sr(1) |
| 1122 | .m(m) |
| 1123 | .n(n) |
| 1124 | .k(k) |
| 1125 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1126 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1127 | } |
| 1128 | } |
| 1129 | } |
| 1130 | } |
| 1131 | |
| 1132 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_div_8) { |
| 1133 | TEST_REQUIRES_ARM_NEON; |
| 1134 | for (size_t k = 16; k <= 80; k += 8) { |
| 1135 | GemmMicrokernelTester() |
| 1136 | .mr(1) |
| 1137 | .nr(8) |
| 1138 | .kr(2) |
| 1139 | .sr(1) |
| 1140 | .m(1) |
| 1141 | .n(8) |
| 1142 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1143 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1144 | } |
| 1145 | } |
| 1146 | |
| 1147 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_div_8_strided_a) { |
| 1148 | TEST_REQUIRES_ARM_NEON; |
| 1149 | for (size_t k = 16; k <= 80; k += 8) { |
| 1150 | GemmMicrokernelTester() |
| 1151 | .mr(1) |
| 1152 | .nr(8) |
| 1153 | .kr(2) |
| 1154 | .sr(1) |
| 1155 | .m(1) |
| 1156 | .n(8) |
| 1157 | .k(k) |
| 1158 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1159 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1160 | } |
| 1161 | } |
| 1162 | |
| 1163 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_div_8_subtile) { |
| 1164 | TEST_REQUIRES_ARM_NEON; |
| 1165 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1166 | for (uint32_t n = 1; n <= 8; n++) { |
| 1167 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1168 | GemmMicrokernelTester() |
| 1169 | .mr(1) |
| 1170 | .nr(8) |
| 1171 | .kr(2) |
| 1172 | .sr(1) |
| 1173 | .m(m) |
| 1174 | .n(n) |
| 1175 | .k(k) |
| 1176 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1177 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1178 | } |
| 1179 | } |
| 1180 | } |
| 1181 | } |
| 1182 | |
| 1183 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8) { |
| 1184 | TEST_REQUIRES_ARM_NEON; |
| 1185 | for (uint32_t n = 9; n < 16; n++) { |
| 1186 | for (size_t k = 1; k <= 40; k += 9) { |
| 1187 | GemmMicrokernelTester() |
| 1188 | .mr(1) |
| 1189 | .nr(8) |
| 1190 | .kr(2) |
| 1191 | .sr(1) |
| 1192 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 1193 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1194 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1195 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1196 | } |
| 1197 | } |
| 1198 | } |
| 1199 | |
| 1200 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_strided_cn) { |
| 1201 | TEST_REQUIRES_ARM_NEON; |
| 1202 | for (uint32_t n = 9; n < 16; n++) { |
| 1203 | for (size_t k = 1; k <= 40; k += 9) { |
| 1204 | GemmMicrokernelTester() |
| 1205 | .mr(1) |
| 1206 | .nr(8) |
| 1207 | .kr(2) |
| 1208 | .sr(1) |
| 1209 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 1210 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1211 | .k(k) |
| 1212 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1213 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1214 | } |
| 1215 | } |
| 1216 | } |
| 1217 | |
| 1218 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_strided_a) { |
| 1219 | TEST_REQUIRES_ARM_NEON; |
| 1220 | for (uint32_t n = 9; n < 16; n++) { |
| 1221 | for (size_t k = 1; k <= 40; k += 9) { |
| 1222 | GemmMicrokernelTester() |
| 1223 | .mr(1) |
| 1224 | .nr(8) |
| 1225 | .kr(2) |
| 1226 | .sr(1) |
| 1227 | .m(1) |
| 1228 | .n(n) |
| 1229 | .k(k) |
| 1230 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1231 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1232 | } |
| 1233 | } |
| 1234 | } |
| 1235 | |
| 1236 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_subtile) { |
| 1237 | TEST_REQUIRES_ARM_NEON; |
| 1238 | for (uint32_t n = 9; n < 16; n++) { |
| 1239 | for (size_t k = 1; k <= 40; k += 9) { |
| 1240 | for (uint32_t m = 1; m <= 1; m++) { |
| 1241 | GemmMicrokernelTester() |
| 1242 | .mr(1) |
| 1243 | .nr(8) |
| 1244 | .kr(2) |
| 1245 | .sr(1) |
| 1246 | .m(m) |
| 1247 | .n(n) |
| 1248 | .k(k) |
| 1249 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1250 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1251 | } |
| 1252 | } |
| 1253 | } |
| 1254 | } |
| 1255 | |
| 1256 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8) { |
| 1257 | TEST_REQUIRES_ARM_NEON; |
| 1258 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1259 | for (size_t k = 1; k <= 40; k += 9) { |
| 1260 | GemmMicrokernelTester() |
| 1261 | .mr(1) |
| 1262 | .nr(8) |
| 1263 | .kr(2) |
| 1264 | .sr(1) |
| 1265 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 1266 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1267 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1268 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1269 | } |
| 1270 | } |
| 1271 | } |
| 1272 | |
| 1273 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_strided_cn) { |
| 1274 | TEST_REQUIRES_ARM_NEON; |
| 1275 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1276 | for (size_t k = 1; k <= 40; k += 9) { |
| 1277 | GemmMicrokernelTester() |
| 1278 | .mr(1) |
| 1279 | .nr(8) |
| 1280 | .kr(2) |
| 1281 | .sr(1) |
| 1282 | .m(1) |
| 1283 | .n(n) |
| 1284 | .k(k) |
| 1285 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1286 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1287 | } |
| 1288 | } |
| 1289 | } |
| 1290 | |
| 1291 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_strided_a) { |
| 1292 | TEST_REQUIRES_ARM_NEON; |
| 1293 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1294 | for (size_t k = 1; k <= 40; k += 9) { |
| 1295 | GemmMicrokernelTester() |
| 1296 | .mr(1) |
| 1297 | .nr(8) |
| 1298 | .kr(2) |
| 1299 | .sr(1) |
| 1300 | .m(1) |
| 1301 | .n(n) |
| 1302 | .k(k) |
| 1303 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1304 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1305 | } |
| 1306 | } |
| 1307 | } |
| 1308 | |
| 1309 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_subtile) { |
| 1310 | TEST_REQUIRES_ARM_NEON; |
| 1311 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1312 | for (size_t k = 1; k <= 40; k += 9) { |
| 1313 | for (uint32_t m = 1; m <= 1; m++) { |
| 1314 | GemmMicrokernelTester() |
| 1315 | .mr(1) |
| 1316 | .nr(8) |
| 1317 | .kr(2) |
| 1318 | .sr(1) |
| 1319 | .m(m) |
| 1320 | .n(n) |
| 1321 | .k(k) |
| 1322 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1323 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1324 | } |
| 1325 | } |
| 1326 | } |
| 1327 | } |
| 1328 | |
| 1329 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, strided_cm_subtile) { |
| 1330 | TEST_REQUIRES_ARM_NEON; |
| 1331 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1332 | for (uint32_t n = 1; n <= 8; n++) { |
| 1333 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1334 | GemmMicrokernelTester() |
| 1335 | .mr(1) |
| 1336 | .nr(8) |
| 1337 | .kr(2) |
| 1338 | .sr(1) |
| 1339 | .m(m) |
| 1340 | .n(n) |
| 1341 | .k(k) |
| 1342 | .cm_stride(11) |
| 1343 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1344 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1345 | } |
| 1346 | } |
| 1347 | } |
| 1348 | } |
| 1349 | |
| 1350 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, qmin) { |
| 1351 | TEST_REQUIRES_ARM_NEON; |
| 1352 | GemmMicrokernelTester() |
| 1353 | .mr(1) |
| 1354 | .nr(8) |
| 1355 | .kr(2) |
| 1356 | .sr(1) |
| 1357 | .m(1) |
| 1358 | .n(8) |
| 1359 | .k(8) |
| 1360 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1361 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1362 | } |
| 1363 | |
| 1364 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, qmax) { |
| 1365 | TEST_REQUIRES_ARM_NEON; |
| 1366 | GemmMicrokernelTester() |
| 1367 | .mr(1) |
| 1368 | .nr(8) |
| 1369 | .kr(2) |
| 1370 | .sr(1) |
| 1371 | .m(1) |
| 1372 | .n(8) |
| 1373 | .k(8) |
| 1374 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1375 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1376 | } |
| 1377 | |
| 1378 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, strided_cm) { |
| 1379 | TEST_REQUIRES_ARM_NEON; |
| 1380 | GemmMicrokernelTester() |
| 1381 | .mr(1) |
| 1382 | .nr(8) |
| 1383 | .kr(2) |
| 1384 | .sr(1) |
| 1385 | .m(1) |
| 1386 | .n(8) |
| 1387 | .k(8) |
| 1388 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1389 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1390 | } |
| 1391 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1392 | |
| 1393 | |
| 1394 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1395 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_eq_8) { |
| 1396 | TEST_REQUIRES_ARM_NEON; |
| 1397 | GemmMicrokernelTester() |
| 1398 | .mr(4) |
| 1399 | .nr(8) |
| 1400 | .kr(2) |
| 1401 | .sr(1) |
| 1402 | .m(4) |
| 1403 | .n(8) |
| 1404 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1405 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1406 | } |
| 1407 | |
| 1408 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, strided_cn) { |
| 1409 | TEST_REQUIRES_ARM_NEON; |
| 1410 | GemmMicrokernelTester() |
| 1411 | .mr(4) |
| 1412 | .nr(8) |
| 1413 | .kr(2) |
| 1414 | .sr(1) |
| 1415 | .m(4) |
| 1416 | .n(8) |
| 1417 | .k(8) |
| 1418 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1419 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1420 | } |
| 1421 | |
| 1422 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_eq_8_strided_a) { |
| 1423 | TEST_REQUIRES_ARM_NEON; |
| 1424 | GemmMicrokernelTester() |
| 1425 | .mr(4) |
| 1426 | .nr(8) |
| 1427 | .kr(2) |
| 1428 | .sr(1) |
| 1429 | .m(4) |
| 1430 | .n(8) |
| 1431 | .k(8) |
| 1432 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1433 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1434 | } |
| 1435 | |
| 1436 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_eq_8_subtile) { |
| 1437 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1438 | for (uint32_t n = 1; n <= 8; n++) { |
| 1439 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1440 | GemmMicrokernelTester() |
| 1441 | .mr(4) |
| 1442 | .nr(8) |
| 1443 | .kr(2) |
| 1444 | .sr(1) |
| 1445 | .m(m) |
| 1446 | .n(n) |
| 1447 | .k(8) |
| 1448 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1449 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1450 | } |
| 1451 | } |
| 1452 | } |
| 1453 | |
| 1454 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_eq_8_subtile_m) { |
| 1455 | TEST_REQUIRES_ARM_NEON; |
| 1456 | for (uint32_t m = 1; m <= 4; m++) { |
| 1457 | GemmMicrokernelTester() |
| 1458 | .mr(4) |
| 1459 | .nr(8) |
| 1460 | .kr(2) |
| 1461 | .sr(1) |
| 1462 | .m(m) |
| 1463 | .n(8) |
| 1464 | .k(8) |
| 1465 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1466 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1467 | } |
| 1468 | } |
| 1469 | |
| 1470 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_eq_8_subtile_n) { |
| 1471 | TEST_REQUIRES_ARM_NEON; |
| 1472 | for (uint32_t n = 1; n <= 8; n++) { |
| 1473 | GemmMicrokernelTester() |
| 1474 | .mr(4) |
| 1475 | .nr(8) |
| 1476 | .kr(2) |
| 1477 | .sr(1) |
| 1478 | .m(4) |
| 1479 | .n(n) |
| 1480 | .k(8) |
| 1481 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1482 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1483 | } |
| 1484 | } |
| 1485 | |
| 1486 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_lt_8) { |
| 1487 | TEST_REQUIRES_ARM_NEON; |
| 1488 | for (size_t k = 1; k < 8; k++) { |
| 1489 | GemmMicrokernelTester() |
| 1490 | .mr(4) |
| 1491 | .nr(8) |
| 1492 | .kr(2) |
| 1493 | .sr(1) |
| 1494 | .m(4) |
| 1495 | .n(8) |
| 1496 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1497 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1498 | } |
| 1499 | } |
| 1500 | |
| 1501 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_lt_8_strided_a) { |
| 1502 | TEST_REQUIRES_ARM_NEON; |
| 1503 | for (size_t k = 1; k < 8; k++) { |
| 1504 | GemmMicrokernelTester() |
| 1505 | .mr(4) |
| 1506 | .nr(8) |
| 1507 | .kr(2) |
| 1508 | .sr(1) |
| 1509 | .m(4) |
| 1510 | .n(8) |
| 1511 | .k(k) |
| 1512 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1513 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1514 | } |
| 1515 | } |
| 1516 | |
| 1517 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_lt_8_subtile) { |
| 1518 | TEST_REQUIRES_ARM_NEON; |
| 1519 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1520 | for (uint32_t n = 1; n <= 8; n++) { |
| 1521 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1522 | GemmMicrokernelTester() |
| 1523 | .mr(4) |
| 1524 | .nr(8) |
| 1525 | .kr(2) |
| 1526 | .sr(1) |
| 1527 | .m(m) |
| 1528 | .n(n) |
| 1529 | .k(k) |
| 1530 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1531 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1532 | } |
| 1533 | } |
| 1534 | } |
| 1535 | } |
| 1536 | |
| 1537 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_gt_8) { |
| 1538 | TEST_REQUIRES_ARM_NEON; |
| 1539 | for (size_t k = 9; k < 16; k++) { |
| 1540 | GemmMicrokernelTester() |
| 1541 | .mr(4) |
| 1542 | .nr(8) |
| 1543 | .kr(2) |
| 1544 | .sr(1) |
| 1545 | .m(4) |
| 1546 | .n(8) |
| 1547 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1548 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1549 | } |
| 1550 | } |
| 1551 | |
| 1552 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_gt_8_strided_a) { |
| 1553 | TEST_REQUIRES_ARM_NEON; |
| 1554 | for (size_t k = 9; k < 16; k++) { |
| 1555 | GemmMicrokernelTester() |
| 1556 | .mr(4) |
| 1557 | .nr(8) |
| 1558 | .kr(2) |
| 1559 | .sr(1) |
| 1560 | .m(4) |
| 1561 | .n(8) |
| 1562 | .k(k) |
| 1563 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1564 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1565 | } |
| 1566 | } |
| 1567 | |
| 1568 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_gt_8_subtile) { |
| 1569 | TEST_REQUIRES_ARM_NEON; |
| 1570 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1571 | for (uint32_t n = 1; n <= 8; n++) { |
| 1572 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1573 | GemmMicrokernelTester() |
| 1574 | .mr(4) |
| 1575 | .nr(8) |
| 1576 | .kr(2) |
| 1577 | .sr(1) |
| 1578 | .m(m) |
| 1579 | .n(n) |
| 1580 | .k(k) |
| 1581 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1582 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1583 | } |
| 1584 | } |
| 1585 | } |
| 1586 | } |
| 1587 | |
| 1588 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_div_8) { |
| 1589 | TEST_REQUIRES_ARM_NEON; |
| 1590 | for (size_t k = 16; k <= 80; k += 8) { |
| 1591 | GemmMicrokernelTester() |
| 1592 | .mr(4) |
| 1593 | .nr(8) |
| 1594 | .kr(2) |
| 1595 | .sr(1) |
| 1596 | .m(4) |
| 1597 | .n(8) |
| 1598 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1599 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1600 | } |
| 1601 | } |
| 1602 | |
| 1603 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_div_8_strided_a) { |
| 1604 | TEST_REQUIRES_ARM_NEON; |
| 1605 | for (size_t k = 16; k <= 80; k += 8) { |
| 1606 | GemmMicrokernelTester() |
| 1607 | .mr(4) |
| 1608 | .nr(8) |
| 1609 | .kr(2) |
| 1610 | .sr(1) |
| 1611 | .m(4) |
| 1612 | .n(8) |
| 1613 | .k(k) |
| 1614 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1615 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1616 | } |
| 1617 | } |
| 1618 | |
| 1619 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, k_div_8_subtile) { |
| 1620 | TEST_REQUIRES_ARM_NEON; |
| 1621 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1622 | for (uint32_t n = 1; n <= 8; n++) { |
| 1623 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1624 | GemmMicrokernelTester() |
| 1625 | .mr(4) |
| 1626 | .nr(8) |
| 1627 | .kr(2) |
| 1628 | .sr(1) |
| 1629 | .m(m) |
| 1630 | .n(n) |
| 1631 | .k(k) |
| 1632 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1633 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1634 | } |
| 1635 | } |
| 1636 | } |
| 1637 | } |
| 1638 | |
| 1639 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_gt_8) { |
| 1640 | TEST_REQUIRES_ARM_NEON; |
| 1641 | for (uint32_t n = 9; n < 16; n++) { |
| 1642 | for (size_t k = 1; k <= 40; k += 9) { |
| 1643 | GemmMicrokernelTester() |
| 1644 | .mr(4) |
| 1645 | .nr(8) |
| 1646 | .kr(2) |
| 1647 | .sr(1) |
| 1648 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 1649 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1650 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1651 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1652 | } |
| 1653 | } |
| 1654 | } |
| 1655 | |
| 1656 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_gt_8_strided_cn) { |
| 1657 | TEST_REQUIRES_ARM_NEON; |
| 1658 | for (uint32_t n = 9; n < 16; n++) { |
| 1659 | for (size_t k = 1; k <= 40; k += 9) { |
| 1660 | GemmMicrokernelTester() |
| 1661 | .mr(4) |
| 1662 | .nr(8) |
| 1663 | .kr(2) |
| 1664 | .sr(1) |
| 1665 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 1666 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1667 | .k(k) |
| 1668 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1669 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1670 | } |
| 1671 | } |
| 1672 | } |
| 1673 | |
| 1674 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_gt_8_strided_a) { |
| 1675 | TEST_REQUIRES_ARM_NEON; |
| 1676 | for (uint32_t n = 9; n < 16; n++) { |
| 1677 | for (size_t k = 1; k <= 40; k += 9) { |
| 1678 | GemmMicrokernelTester() |
| 1679 | .mr(4) |
| 1680 | .nr(8) |
| 1681 | .kr(2) |
| 1682 | .sr(1) |
| 1683 | .m(4) |
| 1684 | .n(n) |
| 1685 | .k(k) |
| 1686 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1687 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1688 | } |
| 1689 | } |
| 1690 | } |
| 1691 | |
| 1692 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_gt_8_subtile) { |
| 1693 | TEST_REQUIRES_ARM_NEON; |
| 1694 | for (uint32_t n = 9; n < 16; n++) { |
| 1695 | for (size_t k = 1; k <= 40; k += 9) { |
| 1696 | for (uint32_t m = 1; m <= 4; m++) { |
| 1697 | GemmMicrokernelTester() |
| 1698 | .mr(4) |
| 1699 | .nr(8) |
| 1700 | .kr(2) |
| 1701 | .sr(1) |
| 1702 | .m(m) |
| 1703 | .n(n) |
| 1704 | .k(k) |
| 1705 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1706 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1707 | } |
| 1708 | } |
| 1709 | } |
| 1710 | } |
| 1711 | |
| 1712 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_div_8) { |
| 1713 | TEST_REQUIRES_ARM_NEON; |
| 1714 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1715 | for (size_t k = 1; k <= 40; k += 9) { |
| 1716 | GemmMicrokernelTester() |
| 1717 | .mr(4) |
| 1718 | .nr(8) |
| 1719 | .kr(2) |
| 1720 | .sr(1) |
| 1721 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 1722 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1723 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1724 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1725 | } |
| 1726 | } |
| 1727 | } |
| 1728 | |
| 1729 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_div_8_strided_cn) { |
| 1730 | TEST_REQUIRES_ARM_NEON; |
| 1731 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1732 | for (size_t k = 1; k <= 40; k += 9) { |
| 1733 | GemmMicrokernelTester() |
| 1734 | .mr(4) |
| 1735 | .nr(8) |
| 1736 | .kr(2) |
| 1737 | .sr(1) |
| 1738 | .m(4) |
| 1739 | .n(n) |
| 1740 | .k(k) |
| 1741 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1742 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1743 | } |
| 1744 | } |
| 1745 | } |
| 1746 | |
| 1747 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_div_8_strided_a) { |
| 1748 | TEST_REQUIRES_ARM_NEON; |
| 1749 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1750 | for (size_t k = 1; k <= 40; k += 9) { |
| 1751 | GemmMicrokernelTester() |
| 1752 | .mr(4) |
| 1753 | .nr(8) |
| 1754 | .kr(2) |
| 1755 | .sr(1) |
| 1756 | .m(4) |
| 1757 | .n(n) |
| 1758 | .k(k) |
| 1759 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1760 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1761 | } |
| 1762 | } |
| 1763 | } |
| 1764 | |
| 1765 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, n_div_8_subtile) { |
| 1766 | TEST_REQUIRES_ARM_NEON; |
| 1767 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1768 | for (size_t k = 1; k <= 40; k += 9) { |
| 1769 | for (uint32_t m = 1; m <= 4; m++) { |
| 1770 | GemmMicrokernelTester() |
| 1771 | .mr(4) |
| 1772 | .nr(8) |
| 1773 | .kr(2) |
| 1774 | .sr(1) |
| 1775 | .m(m) |
| 1776 | .n(n) |
| 1777 | .k(k) |
| 1778 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1779 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1780 | } |
| 1781 | } |
| 1782 | } |
| 1783 | } |
| 1784 | |
| 1785 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, strided_cm_subtile) { |
| 1786 | TEST_REQUIRES_ARM_NEON; |
| 1787 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1788 | for (uint32_t n = 1; n <= 8; n++) { |
| 1789 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1790 | GemmMicrokernelTester() |
| 1791 | .mr(4) |
| 1792 | .nr(8) |
| 1793 | .kr(2) |
| 1794 | .sr(1) |
| 1795 | .m(m) |
| 1796 | .n(n) |
| 1797 | .k(k) |
| 1798 | .cm_stride(11) |
| 1799 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1800 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1801 | } |
| 1802 | } |
| 1803 | } |
| 1804 | } |
| 1805 | |
| 1806 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, qmin) { |
| 1807 | TEST_REQUIRES_ARM_NEON; |
| 1808 | GemmMicrokernelTester() |
| 1809 | .mr(4) |
| 1810 | .nr(8) |
| 1811 | .kr(2) |
| 1812 | .sr(1) |
| 1813 | .m(4) |
| 1814 | .n(8) |
| 1815 | .k(8) |
| 1816 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1817 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1818 | } |
| 1819 | |
| 1820 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, qmax) { |
| 1821 | TEST_REQUIRES_ARM_NEON; |
| 1822 | GemmMicrokernelTester() |
| 1823 | .mr(4) |
| 1824 | .nr(8) |
| 1825 | .kr(2) |
| 1826 | .sr(1) |
| 1827 | .m(4) |
| 1828 | .n(8) |
| 1829 | .k(8) |
| 1830 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1831 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1832 | } |
| 1833 | |
| 1834 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD1R, strided_cm) { |
| 1835 | TEST_REQUIRES_ARM_NEON; |
| 1836 | GemmMicrokernelTester() |
| 1837 | .mr(4) |
| 1838 | .nr(8) |
| 1839 | .kr(2) |
| 1840 | .sr(1) |
| 1841 | .m(4) |
| 1842 | .n(8) |
| 1843 | .k(8) |
| 1844 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1845 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1846 | } |
| 1847 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1848 | |
| 1849 | |
| 1850 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 1851 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_eq_8) { |
| 1852 | TEST_REQUIRES_ARM_NEON; |
| 1853 | GemmMicrokernelTester() |
| 1854 | .mr(1) |
| 1855 | .nr(16) |
| 1856 | .kr(2) |
| 1857 | .sr(1) |
| 1858 | .m(1) |
| 1859 | .n(16) |
| 1860 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1861 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1862 | } |
| 1863 | |
| 1864 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, strided_cn) { |
| 1865 | TEST_REQUIRES_ARM_NEON; |
| 1866 | GemmMicrokernelTester() |
| 1867 | .mr(1) |
| 1868 | .nr(16) |
| 1869 | .kr(2) |
| 1870 | .sr(1) |
| 1871 | .m(1) |
| 1872 | .n(16) |
| 1873 | .k(8) |
| 1874 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1875 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1876 | } |
| 1877 | |
| 1878 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_eq_8_strided_a) { |
| 1879 | TEST_REQUIRES_ARM_NEON; |
| 1880 | GemmMicrokernelTester() |
| 1881 | .mr(1) |
| 1882 | .nr(16) |
| 1883 | .kr(2) |
| 1884 | .sr(1) |
| 1885 | .m(1) |
| 1886 | .n(16) |
| 1887 | .k(8) |
| 1888 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1889 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1890 | } |
| 1891 | |
| 1892 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_eq_8_subtile) { |
| 1893 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1894 | for (uint32_t n = 1; n <= 16; n++) { |
| 1895 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1896 | GemmMicrokernelTester() |
| 1897 | .mr(1) |
| 1898 | .nr(16) |
| 1899 | .kr(2) |
| 1900 | .sr(1) |
| 1901 | .m(m) |
| 1902 | .n(n) |
| 1903 | .k(8) |
| 1904 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1905 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1906 | } |
| 1907 | } |
| 1908 | } |
| 1909 | |
| 1910 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_eq_8_subtile_m) { |
| 1911 | TEST_REQUIRES_ARM_NEON; |
| 1912 | for (uint32_t m = 1; m <= 1; m++) { |
| 1913 | GemmMicrokernelTester() |
| 1914 | .mr(1) |
| 1915 | .nr(16) |
| 1916 | .kr(2) |
| 1917 | .sr(1) |
| 1918 | .m(m) |
| 1919 | .n(16) |
| 1920 | .k(8) |
| 1921 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1922 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1923 | } |
| 1924 | } |
| 1925 | |
| 1926 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_eq_8_subtile_n) { |
| 1927 | TEST_REQUIRES_ARM_NEON; |
| 1928 | for (uint32_t n = 1; n <= 16; n++) { |
| 1929 | GemmMicrokernelTester() |
| 1930 | .mr(1) |
| 1931 | .nr(16) |
| 1932 | .kr(2) |
| 1933 | .sr(1) |
| 1934 | .m(1) |
| 1935 | .n(n) |
| 1936 | .k(8) |
| 1937 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1938 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1939 | } |
| 1940 | } |
| 1941 | |
| 1942 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_lt_8) { |
| 1943 | TEST_REQUIRES_ARM_NEON; |
| 1944 | for (size_t k = 1; k < 8; k++) { |
| 1945 | GemmMicrokernelTester() |
| 1946 | .mr(1) |
| 1947 | .nr(16) |
| 1948 | .kr(2) |
| 1949 | .sr(1) |
| 1950 | .m(1) |
| 1951 | .n(16) |
| 1952 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1953 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1954 | } |
| 1955 | } |
| 1956 | |
| 1957 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_lt_8_strided_a) { |
| 1958 | TEST_REQUIRES_ARM_NEON; |
| 1959 | for (size_t k = 1; k < 8; k++) { |
| 1960 | GemmMicrokernelTester() |
| 1961 | .mr(1) |
| 1962 | .nr(16) |
| 1963 | .kr(2) |
| 1964 | .sr(1) |
| 1965 | .m(1) |
| 1966 | .n(16) |
| 1967 | .k(k) |
| 1968 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1969 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1970 | } |
| 1971 | } |
| 1972 | |
| 1973 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_lt_8_subtile) { |
| 1974 | TEST_REQUIRES_ARM_NEON; |
| 1975 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 1976 | for (uint32_t n = 1; n <= 16; n++) { |
| 1977 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1978 | GemmMicrokernelTester() |
| 1979 | .mr(1) |
| 1980 | .nr(16) |
| 1981 | .kr(2) |
| 1982 | .sr(1) |
| 1983 | .m(m) |
| 1984 | .n(n) |
| 1985 | .k(k) |
| 1986 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 1987 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 1988 | } |
| 1989 | } |
| 1990 | } |
| 1991 | } |
| 1992 | |
| 1993 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_gt_8) { |
| 1994 | TEST_REQUIRES_ARM_NEON; |
| 1995 | for (size_t k = 9; k < 16; k++) { |
| 1996 | GemmMicrokernelTester() |
| 1997 | .mr(1) |
| 1998 | .nr(16) |
| 1999 | .kr(2) |
| 2000 | .sr(1) |
| 2001 | .m(1) |
| 2002 | .n(16) |
| 2003 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2004 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2005 | } |
| 2006 | } |
| 2007 | |
| 2008 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_gt_8_strided_a) { |
| 2009 | TEST_REQUIRES_ARM_NEON; |
| 2010 | for (size_t k = 9; k < 16; k++) { |
| 2011 | GemmMicrokernelTester() |
| 2012 | .mr(1) |
| 2013 | .nr(16) |
| 2014 | .kr(2) |
| 2015 | .sr(1) |
| 2016 | .m(1) |
| 2017 | .n(16) |
| 2018 | .k(k) |
| 2019 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2020 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2021 | } |
| 2022 | } |
| 2023 | |
| 2024 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_gt_8_subtile) { |
| 2025 | TEST_REQUIRES_ARM_NEON; |
| 2026 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2027 | for (uint32_t n = 1; n <= 16; n++) { |
| 2028 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2029 | GemmMicrokernelTester() |
| 2030 | .mr(1) |
| 2031 | .nr(16) |
| 2032 | .kr(2) |
| 2033 | .sr(1) |
| 2034 | .m(m) |
| 2035 | .n(n) |
| 2036 | .k(k) |
| 2037 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2038 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2039 | } |
| 2040 | } |
| 2041 | } |
| 2042 | } |
| 2043 | |
| 2044 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_div_8) { |
| 2045 | TEST_REQUIRES_ARM_NEON; |
| 2046 | for (size_t k = 16; k <= 80; k += 8) { |
| 2047 | GemmMicrokernelTester() |
| 2048 | .mr(1) |
| 2049 | .nr(16) |
| 2050 | .kr(2) |
| 2051 | .sr(1) |
| 2052 | .m(1) |
| 2053 | .n(16) |
| 2054 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2055 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2056 | } |
| 2057 | } |
| 2058 | |
| 2059 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_div_8_strided_a) { |
| 2060 | TEST_REQUIRES_ARM_NEON; |
| 2061 | for (size_t k = 16; k <= 80; k += 8) { |
| 2062 | GemmMicrokernelTester() |
| 2063 | .mr(1) |
| 2064 | .nr(16) |
| 2065 | .kr(2) |
| 2066 | .sr(1) |
| 2067 | .m(1) |
| 2068 | .n(16) |
| 2069 | .k(k) |
| 2070 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2071 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2072 | } |
| 2073 | } |
| 2074 | |
| 2075 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, k_div_8_subtile) { |
| 2076 | TEST_REQUIRES_ARM_NEON; |
| 2077 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2078 | for (uint32_t n = 1; n <= 16; n++) { |
| 2079 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2080 | GemmMicrokernelTester() |
| 2081 | .mr(1) |
| 2082 | .nr(16) |
| 2083 | .kr(2) |
| 2084 | .sr(1) |
| 2085 | .m(m) |
| 2086 | .n(n) |
| 2087 | .k(k) |
| 2088 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2089 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2090 | } |
| 2091 | } |
| 2092 | } |
| 2093 | } |
| 2094 | |
| 2095 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_gt_16) { |
| 2096 | TEST_REQUIRES_ARM_NEON; |
| 2097 | for (uint32_t n = 17; n < 32; n++) { |
| 2098 | for (size_t k = 1; k <= 40; k += 9) { |
| 2099 | GemmMicrokernelTester() |
| 2100 | .mr(1) |
| 2101 | .nr(16) |
| 2102 | .kr(2) |
| 2103 | .sr(1) |
| 2104 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2105 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2106 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2107 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2108 | } |
| 2109 | } |
| 2110 | } |
| 2111 | |
| 2112 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_gt_16_strided_cn) { |
| 2113 | TEST_REQUIRES_ARM_NEON; |
| 2114 | for (uint32_t n = 17; n < 32; n++) { |
| 2115 | for (size_t k = 1; k <= 40; k += 9) { |
| 2116 | GemmMicrokernelTester() |
| 2117 | .mr(1) |
| 2118 | .nr(16) |
| 2119 | .kr(2) |
| 2120 | .sr(1) |
| 2121 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2122 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2123 | .k(k) |
| 2124 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2125 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2126 | } |
| 2127 | } |
| 2128 | } |
| 2129 | |
| 2130 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_gt_16_strided_a) { |
| 2131 | TEST_REQUIRES_ARM_NEON; |
| 2132 | for (uint32_t n = 17; n < 32; n++) { |
| 2133 | for (size_t k = 1; k <= 40; k += 9) { |
| 2134 | GemmMicrokernelTester() |
| 2135 | .mr(1) |
| 2136 | .nr(16) |
| 2137 | .kr(2) |
| 2138 | .sr(1) |
| 2139 | .m(1) |
| 2140 | .n(n) |
| 2141 | .k(k) |
| 2142 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2143 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2144 | } |
| 2145 | } |
| 2146 | } |
| 2147 | |
| 2148 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_gt_16_subtile) { |
| 2149 | TEST_REQUIRES_ARM_NEON; |
| 2150 | for (uint32_t n = 17; n < 32; n++) { |
| 2151 | for (size_t k = 1; k <= 40; k += 9) { |
| 2152 | for (uint32_t m = 1; m <= 1; m++) { |
| 2153 | GemmMicrokernelTester() |
| 2154 | .mr(1) |
| 2155 | .nr(16) |
| 2156 | .kr(2) |
| 2157 | .sr(1) |
| 2158 | .m(m) |
| 2159 | .n(n) |
| 2160 | .k(k) |
| 2161 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2162 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2163 | } |
| 2164 | } |
| 2165 | } |
| 2166 | } |
| 2167 | |
| 2168 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_div_16) { |
| 2169 | TEST_REQUIRES_ARM_NEON; |
| 2170 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2171 | for (size_t k = 1; k <= 40; k += 9) { |
| 2172 | GemmMicrokernelTester() |
| 2173 | .mr(1) |
| 2174 | .nr(16) |
| 2175 | .kr(2) |
| 2176 | .sr(1) |
| 2177 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2178 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2179 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2180 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2181 | } |
| 2182 | } |
| 2183 | } |
| 2184 | |
| 2185 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_div_16_strided_cn) { |
| 2186 | TEST_REQUIRES_ARM_NEON; |
| 2187 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2188 | for (size_t k = 1; k <= 40; k += 9) { |
| 2189 | GemmMicrokernelTester() |
| 2190 | .mr(1) |
| 2191 | .nr(16) |
| 2192 | .kr(2) |
| 2193 | .sr(1) |
| 2194 | .m(1) |
| 2195 | .n(n) |
| 2196 | .k(k) |
| 2197 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2198 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2199 | } |
| 2200 | } |
| 2201 | } |
| 2202 | |
| 2203 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_div_16_strided_a) { |
| 2204 | TEST_REQUIRES_ARM_NEON; |
| 2205 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2206 | for (size_t k = 1; k <= 40; k += 9) { |
| 2207 | GemmMicrokernelTester() |
| 2208 | .mr(1) |
| 2209 | .nr(16) |
| 2210 | .kr(2) |
| 2211 | .sr(1) |
| 2212 | .m(1) |
| 2213 | .n(n) |
| 2214 | .k(k) |
| 2215 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2216 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2217 | } |
| 2218 | } |
| 2219 | } |
| 2220 | |
| 2221 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, n_div_16_subtile) { |
| 2222 | TEST_REQUIRES_ARM_NEON; |
| 2223 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2224 | for (size_t k = 1; k <= 40; k += 9) { |
| 2225 | for (uint32_t m = 1; m <= 1; m++) { |
| 2226 | GemmMicrokernelTester() |
| 2227 | .mr(1) |
| 2228 | .nr(16) |
| 2229 | .kr(2) |
| 2230 | .sr(1) |
| 2231 | .m(m) |
| 2232 | .n(n) |
| 2233 | .k(k) |
| 2234 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2235 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2236 | } |
| 2237 | } |
| 2238 | } |
| 2239 | } |
| 2240 | |
| 2241 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, strided_cm_subtile) { |
| 2242 | TEST_REQUIRES_ARM_NEON; |
| 2243 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2244 | for (uint32_t n = 1; n <= 16; n++) { |
| 2245 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2246 | GemmMicrokernelTester() |
| 2247 | .mr(1) |
| 2248 | .nr(16) |
| 2249 | .kr(2) |
| 2250 | .sr(1) |
| 2251 | .m(m) |
| 2252 | .n(n) |
| 2253 | .k(k) |
| 2254 | .cm_stride(19) |
| 2255 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2256 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2257 | } |
| 2258 | } |
| 2259 | } |
| 2260 | } |
| 2261 | |
| 2262 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, qmin) { |
| 2263 | TEST_REQUIRES_ARM_NEON; |
| 2264 | GemmMicrokernelTester() |
| 2265 | .mr(1) |
| 2266 | .nr(16) |
| 2267 | .kr(2) |
| 2268 | .sr(1) |
| 2269 | .m(1) |
| 2270 | .n(16) |
| 2271 | .k(8) |
| 2272 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2273 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2274 | } |
| 2275 | |
| 2276 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, qmax) { |
| 2277 | TEST_REQUIRES_ARM_NEON; |
| 2278 | GemmMicrokernelTester() |
| 2279 | .mr(1) |
| 2280 | .nr(16) |
| 2281 | .kr(2) |
| 2282 | .sr(1) |
| 2283 | .m(1) |
| 2284 | .n(16) |
| 2285 | .k(8) |
| 2286 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2287 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2288 | } |
| 2289 | |
| 2290 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MULL_LD1R, strided_cm) { |
| 2291 | TEST_REQUIRES_ARM_NEON; |
| 2292 | GemmMicrokernelTester() |
| 2293 | .mr(1) |
| 2294 | .nr(16) |
| 2295 | .kr(2) |
| 2296 | .sr(1) |
| 2297 | .m(1) |
| 2298 | .n(16) |
| 2299 | .k(8) |
| 2300 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2301 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2302 | } |
| 2303 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 2304 | |
| 2305 | |
| 2306 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 2307 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_eq_16) { |
| 2308 | TEST_REQUIRES_ARM_NEON; |
| 2309 | GemmMicrokernelTester() |
| 2310 | .mr(4) |
| 2311 | .nr(16) |
| 2312 | .kr(2) |
| 2313 | .sr(1) |
| 2314 | .m(4) |
| 2315 | .n(16) |
| 2316 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2317 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2318 | } |
| 2319 | |
| 2320 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, strided_cn) { |
| 2321 | TEST_REQUIRES_ARM_NEON; |
| 2322 | GemmMicrokernelTester() |
| 2323 | .mr(4) |
| 2324 | .nr(16) |
| 2325 | .kr(2) |
| 2326 | .sr(1) |
| 2327 | .m(4) |
| 2328 | .n(16) |
| 2329 | .k(16) |
| 2330 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2331 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2332 | } |
| 2333 | |
| 2334 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_eq_16_strided_a) { |
| 2335 | TEST_REQUIRES_ARM_NEON; |
| 2336 | GemmMicrokernelTester() |
| 2337 | .mr(4) |
| 2338 | .nr(16) |
| 2339 | .kr(2) |
| 2340 | .sr(1) |
| 2341 | .m(4) |
| 2342 | .n(16) |
| 2343 | .k(16) |
| 2344 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2345 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2346 | } |
| 2347 | |
| 2348 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_eq_16_subtile) { |
| 2349 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2350 | for (uint32_t n = 1; n <= 16; n++) { |
| 2351 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2352 | GemmMicrokernelTester() |
| 2353 | .mr(4) |
| 2354 | .nr(16) |
| 2355 | .kr(2) |
| 2356 | .sr(1) |
| 2357 | .m(m) |
| 2358 | .n(n) |
| 2359 | .k(16) |
| 2360 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2361 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2362 | } |
| 2363 | } |
| 2364 | } |
| 2365 | |
| 2366 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_eq_16_subtile_m) { |
| 2367 | TEST_REQUIRES_ARM_NEON; |
| 2368 | for (uint32_t m = 1; m <= 4; m++) { |
| 2369 | GemmMicrokernelTester() |
| 2370 | .mr(4) |
| 2371 | .nr(16) |
| 2372 | .kr(2) |
| 2373 | .sr(1) |
| 2374 | .m(m) |
| 2375 | .n(16) |
| 2376 | .k(16) |
| 2377 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2378 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2379 | } |
| 2380 | } |
| 2381 | |
| 2382 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_eq_16_subtile_n) { |
| 2383 | TEST_REQUIRES_ARM_NEON; |
| 2384 | for (uint32_t n = 1; n <= 16; n++) { |
| 2385 | GemmMicrokernelTester() |
| 2386 | .mr(4) |
| 2387 | .nr(16) |
| 2388 | .kr(2) |
| 2389 | .sr(1) |
| 2390 | .m(4) |
| 2391 | .n(n) |
| 2392 | .k(16) |
| 2393 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2394 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2395 | } |
| 2396 | } |
| 2397 | |
| 2398 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_lt_16) { |
| 2399 | TEST_REQUIRES_ARM_NEON; |
| 2400 | for (size_t k = 1; k < 16; k++) { |
| 2401 | GemmMicrokernelTester() |
| 2402 | .mr(4) |
| 2403 | .nr(16) |
| 2404 | .kr(2) |
| 2405 | .sr(1) |
| 2406 | .m(4) |
| 2407 | .n(16) |
| 2408 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2409 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2410 | } |
| 2411 | } |
| 2412 | |
| 2413 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_lt_16_strided_a) { |
| 2414 | TEST_REQUIRES_ARM_NEON; |
| 2415 | for (size_t k = 1; k < 16; k++) { |
| 2416 | GemmMicrokernelTester() |
| 2417 | .mr(4) |
| 2418 | .nr(16) |
| 2419 | .kr(2) |
| 2420 | .sr(1) |
| 2421 | .m(4) |
| 2422 | .n(16) |
| 2423 | .k(k) |
| 2424 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2425 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2426 | } |
| 2427 | } |
| 2428 | |
| 2429 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_lt_16_subtile) { |
| 2430 | TEST_REQUIRES_ARM_NEON; |
| 2431 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2432 | for (uint32_t n = 1; n <= 16; n++) { |
| 2433 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2434 | GemmMicrokernelTester() |
| 2435 | .mr(4) |
| 2436 | .nr(16) |
| 2437 | .kr(2) |
| 2438 | .sr(1) |
| 2439 | .m(m) |
| 2440 | .n(n) |
| 2441 | .k(k) |
| 2442 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2443 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2444 | } |
| 2445 | } |
| 2446 | } |
| 2447 | } |
| 2448 | |
| 2449 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_gt_16) { |
| 2450 | TEST_REQUIRES_ARM_NEON; |
| 2451 | for (size_t k = 17; k < 32; k++) { |
| 2452 | GemmMicrokernelTester() |
| 2453 | .mr(4) |
| 2454 | .nr(16) |
| 2455 | .kr(2) |
| 2456 | .sr(1) |
| 2457 | .m(4) |
| 2458 | .n(16) |
| 2459 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2460 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2461 | } |
| 2462 | } |
| 2463 | |
| 2464 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_gt_16_strided_a) { |
| 2465 | TEST_REQUIRES_ARM_NEON; |
| 2466 | for (size_t k = 17; k < 32; k++) { |
| 2467 | GemmMicrokernelTester() |
| 2468 | .mr(4) |
| 2469 | .nr(16) |
| 2470 | .kr(2) |
| 2471 | .sr(1) |
| 2472 | .m(4) |
| 2473 | .n(16) |
| 2474 | .k(k) |
| 2475 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2476 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2477 | } |
| 2478 | } |
| 2479 | |
| 2480 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_gt_16_subtile) { |
| 2481 | TEST_REQUIRES_ARM_NEON; |
| 2482 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2483 | for (uint32_t n = 1; n <= 16; n++) { |
| 2484 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2485 | GemmMicrokernelTester() |
| 2486 | .mr(4) |
| 2487 | .nr(16) |
| 2488 | .kr(2) |
| 2489 | .sr(1) |
| 2490 | .m(m) |
| 2491 | .n(n) |
| 2492 | .k(k) |
| 2493 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2494 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2495 | } |
| 2496 | } |
| 2497 | } |
| 2498 | } |
| 2499 | |
| 2500 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_div_16) { |
| 2501 | TEST_REQUIRES_ARM_NEON; |
| 2502 | for (size_t k = 32; k <= 160; k += 16) { |
| 2503 | GemmMicrokernelTester() |
| 2504 | .mr(4) |
| 2505 | .nr(16) |
| 2506 | .kr(2) |
| 2507 | .sr(1) |
| 2508 | .m(4) |
| 2509 | .n(16) |
| 2510 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2511 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2512 | } |
| 2513 | } |
| 2514 | |
| 2515 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_div_16_strided_a) { |
| 2516 | TEST_REQUIRES_ARM_NEON; |
| 2517 | for (size_t k = 32; k <= 160; k += 16) { |
| 2518 | GemmMicrokernelTester() |
| 2519 | .mr(4) |
| 2520 | .nr(16) |
| 2521 | .kr(2) |
| 2522 | .sr(1) |
| 2523 | .m(4) |
| 2524 | .n(16) |
| 2525 | .k(k) |
| 2526 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2527 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2528 | } |
| 2529 | } |
| 2530 | |
| 2531 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, k_div_16_subtile) { |
| 2532 | TEST_REQUIRES_ARM_NEON; |
| 2533 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2534 | for (uint32_t n = 1; n <= 16; n++) { |
| 2535 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2536 | GemmMicrokernelTester() |
| 2537 | .mr(4) |
| 2538 | .nr(16) |
| 2539 | .kr(2) |
| 2540 | .sr(1) |
| 2541 | .m(m) |
| 2542 | .n(n) |
| 2543 | .k(k) |
| 2544 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2545 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2546 | } |
| 2547 | } |
| 2548 | } |
| 2549 | } |
| 2550 | |
| 2551 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_gt_16) { |
| 2552 | TEST_REQUIRES_ARM_NEON; |
| 2553 | for (uint32_t n = 17; n < 32; n++) { |
| 2554 | for (size_t k = 1; k <= 80; k += 17) { |
| 2555 | GemmMicrokernelTester() |
| 2556 | .mr(4) |
| 2557 | .nr(16) |
| 2558 | .kr(2) |
| 2559 | .sr(1) |
| 2560 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2561 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2562 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2563 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2564 | } |
| 2565 | } |
| 2566 | } |
| 2567 | |
| 2568 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_gt_16_strided_cn) { |
| 2569 | TEST_REQUIRES_ARM_NEON; |
| 2570 | for (uint32_t n = 17; n < 32; n++) { |
| 2571 | for (size_t k = 1; k <= 80; k += 17) { |
| 2572 | GemmMicrokernelTester() |
| 2573 | .mr(4) |
| 2574 | .nr(16) |
| 2575 | .kr(2) |
| 2576 | .sr(1) |
| 2577 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2578 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2579 | .k(k) |
| 2580 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2581 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2582 | } |
| 2583 | } |
| 2584 | } |
| 2585 | |
| 2586 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_gt_16_strided_a) { |
| 2587 | TEST_REQUIRES_ARM_NEON; |
| 2588 | for (uint32_t n = 17; n < 32; n++) { |
| 2589 | for (size_t k = 1; k <= 80; k += 17) { |
| 2590 | GemmMicrokernelTester() |
| 2591 | .mr(4) |
| 2592 | .nr(16) |
| 2593 | .kr(2) |
| 2594 | .sr(1) |
| 2595 | .m(4) |
| 2596 | .n(n) |
| 2597 | .k(k) |
| 2598 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2599 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2600 | } |
| 2601 | } |
| 2602 | } |
| 2603 | |
| 2604 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_gt_16_subtile) { |
| 2605 | TEST_REQUIRES_ARM_NEON; |
| 2606 | for (uint32_t n = 17; n < 32; n++) { |
| 2607 | for (size_t k = 1; k <= 80; k += 17) { |
| 2608 | for (uint32_t m = 1; m <= 4; m++) { |
| 2609 | GemmMicrokernelTester() |
| 2610 | .mr(4) |
| 2611 | .nr(16) |
| 2612 | .kr(2) |
| 2613 | .sr(1) |
| 2614 | .m(m) |
| 2615 | .n(n) |
| 2616 | .k(k) |
| 2617 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2618 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2619 | } |
| 2620 | } |
| 2621 | } |
| 2622 | } |
| 2623 | |
| 2624 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_div_16) { |
| 2625 | TEST_REQUIRES_ARM_NEON; |
| 2626 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2627 | for (size_t k = 1; k <= 80; k += 17) { |
| 2628 | GemmMicrokernelTester() |
| 2629 | .mr(4) |
| 2630 | .nr(16) |
| 2631 | .kr(2) |
| 2632 | .sr(1) |
| 2633 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 2634 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2635 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2636 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2637 | } |
| 2638 | } |
| 2639 | } |
| 2640 | |
| 2641 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_div_16_strided_cn) { |
| 2642 | TEST_REQUIRES_ARM_NEON; |
| 2643 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2644 | for (size_t k = 1; k <= 80; k += 17) { |
| 2645 | GemmMicrokernelTester() |
| 2646 | .mr(4) |
| 2647 | .nr(16) |
| 2648 | .kr(2) |
| 2649 | .sr(1) |
| 2650 | .m(4) |
| 2651 | .n(n) |
| 2652 | .k(k) |
| 2653 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2654 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2655 | } |
| 2656 | } |
| 2657 | } |
| 2658 | |
| 2659 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_div_16_strided_a) { |
| 2660 | TEST_REQUIRES_ARM_NEON; |
| 2661 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2662 | for (size_t k = 1; k <= 80; k += 17) { |
| 2663 | GemmMicrokernelTester() |
| 2664 | .mr(4) |
| 2665 | .nr(16) |
| 2666 | .kr(2) |
| 2667 | .sr(1) |
| 2668 | .m(4) |
| 2669 | .n(n) |
| 2670 | .k(k) |
| 2671 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2672 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2673 | } |
| 2674 | } |
| 2675 | } |
| 2676 | |
| 2677 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, n_div_16_subtile) { |
| 2678 | TEST_REQUIRES_ARM_NEON; |
| 2679 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2680 | for (size_t k = 1; k <= 80; k += 17) { |
| 2681 | for (uint32_t m = 1; m <= 4; m++) { |
| 2682 | GemmMicrokernelTester() |
| 2683 | .mr(4) |
| 2684 | .nr(16) |
| 2685 | .kr(2) |
| 2686 | .sr(1) |
| 2687 | .m(m) |
| 2688 | .n(n) |
| 2689 | .k(k) |
| 2690 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2691 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2692 | } |
| 2693 | } |
| 2694 | } |
| 2695 | } |
| 2696 | |
| 2697 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, strided_cm_subtile) { |
| 2698 | TEST_REQUIRES_ARM_NEON; |
| 2699 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2700 | for (uint32_t n = 1; n <= 16; n++) { |
| 2701 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2702 | GemmMicrokernelTester() |
| 2703 | .mr(4) |
| 2704 | .nr(16) |
| 2705 | .kr(2) |
| 2706 | .sr(1) |
| 2707 | .m(m) |
| 2708 | .n(n) |
| 2709 | .k(k) |
| 2710 | .cm_stride(19) |
| 2711 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2712 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2713 | } |
| 2714 | } |
| 2715 | } |
| 2716 | } |
| 2717 | |
| 2718 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, qmin) { |
| 2719 | TEST_REQUIRES_ARM_NEON; |
| 2720 | GemmMicrokernelTester() |
| 2721 | .mr(4) |
| 2722 | .nr(16) |
| 2723 | .kr(2) |
| 2724 | .sr(1) |
| 2725 | .m(4) |
| 2726 | .n(16) |
| 2727 | .k(16) |
| 2728 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2729 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2730 | } |
| 2731 | |
| 2732 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, qmax) { |
| 2733 | TEST_REQUIRES_ARM_NEON; |
| 2734 | GemmMicrokernelTester() |
| 2735 | .mr(4) |
| 2736 | .nr(16) |
| 2737 | .kr(2) |
| 2738 | .sr(1) |
| 2739 | .m(4) |
| 2740 | .n(16) |
| 2741 | .k(16) |
| 2742 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2743 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2744 | } |
| 2745 | |
| 2746 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MLAL_LD1R, strided_cm) { |
| 2747 | TEST_REQUIRES_ARM_NEON; |
| 2748 | GemmMicrokernelTester() |
| 2749 | .mr(4) |
| 2750 | .nr(16) |
| 2751 | .kr(2) |
| 2752 | .sr(1) |
| 2753 | .m(4) |
| 2754 | .n(16) |
| 2755 | .k(16) |
| 2756 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2757 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2758 | } |
| 2759 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 2760 | |
| 2761 | |
| 2762 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 2763 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8) { |
| 2764 | TEST_REQUIRES_ARM_NEON; |
| 2765 | GemmMicrokernelTester() |
| 2766 | .mr(2) |
| 2767 | .nr(8) |
| 2768 | .kr(2) |
| 2769 | .sr(1) |
| 2770 | .m(2) |
| 2771 | .n(8) |
| 2772 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2773 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2774 | } |
| 2775 | |
| 2776 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, strided_cn) { |
| 2777 | TEST_REQUIRES_ARM_NEON; |
| 2778 | GemmMicrokernelTester() |
| 2779 | .mr(2) |
| 2780 | .nr(8) |
| 2781 | .kr(2) |
| 2782 | .sr(1) |
| 2783 | .m(2) |
| 2784 | .n(8) |
| 2785 | .k(8) |
| 2786 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2787 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2788 | } |
| 2789 | |
| 2790 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8_strided_a) { |
| 2791 | TEST_REQUIRES_ARM_NEON; |
| 2792 | GemmMicrokernelTester() |
| 2793 | .mr(2) |
| 2794 | .nr(8) |
| 2795 | .kr(2) |
| 2796 | .sr(1) |
| 2797 | .m(2) |
| 2798 | .n(8) |
| 2799 | .k(8) |
| 2800 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2801 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2802 | } |
| 2803 | |
| 2804 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8_subtile) { |
| 2805 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2806 | for (uint32_t n = 1; n <= 8; n++) { |
| 2807 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2808 | GemmMicrokernelTester() |
| 2809 | .mr(2) |
| 2810 | .nr(8) |
| 2811 | .kr(2) |
| 2812 | .sr(1) |
| 2813 | .m(m) |
| 2814 | .n(n) |
| 2815 | .k(8) |
| 2816 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2817 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2818 | } |
| 2819 | } |
| 2820 | } |
| 2821 | |
| 2822 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8_subtile_m) { |
| 2823 | TEST_REQUIRES_ARM_NEON; |
| 2824 | for (uint32_t m = 1; m <= 2; m++) { |
| 2825 | GemmMicrokernelTester() |
| 2826 | .mr(2) |
| 2827 | .nr(8) |
| 2828 | .kr(2) |
| 2829 | .sr(1) |
| 2830 | .m(m) |
| 2831 | .n(8) |
| 2832 | .k(8) |
| 2833 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2834 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2835 | } |
| 2836 | } |
| 2837 | |
| 2838 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8_subtile_n) { |
| 2839 | TEST_REQUIRES_ARM_NEON; |
| 2840 | for (uint32_t n = 1; n <= 8; n++) { |
| 2841 | GemmMicrokernelTester() |
| 2842 | .mr(2) |
| 2843 | .nr(8) |
| 2844 | .kr(2) |
| 2845 | .sr(1) |
| 2846 | .m(2) |
| 2847 | .n(n) |
| 2848 | .k(8) |
| 2849 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2850 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2851 | } |
| 2852 | } |
| 2853 | |
| 2854 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_lt_8) { |
| 2855 | TEST_REQUIRES_ARM_NEON; |
| 2856 | for (size_t k = 1; k < 8; k++) { |
| 2857 | GemmMicrokernelTester() |
| 2858 | .mr(2) |
| 2859 | .nr(8) |
| 2860 | .kr(2) |
| 2861 | .sr(1) |
| 2862 | .m(2) |
| 2863 | .n(8) |
| 2864 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2865 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2866 | } |
| 2867 | } |
| 2868 | |
| 2869 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_lt_8_strided_a) { |
| 2870 | TEST_REQUIRES_ARM_NEON; |
| 2871 | for (size_t k = 1; k < 8; k++) { |
| 2872 | GemmMicrokernelTester() |
| 2873 | .mr(2) |
| 2874 | .nr(8) |
| 2875 | .kr(2) |
| 2876 | .sr(1) |
| 2877 | .m(2) |
| 2878 | .n(8) |
| 2879 | .k(k) |
| 2880 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2881 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2882 | } |
| 2883 | } |
| 2884 | |
| 2885 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_lt_8_subtile) { |
| 2886 | TEST_REQUIRES_ARM_NEON; |
| 2887 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2888 | for (uint32_t n = 1; n <= 8; n++) { |
| 2889 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2890 | GemmMicrokernelTester() |
| 2891 | .mr(2) |
| 2892 | .nr(8) |
| 2893 | .kr(2) |
| 2894 | .sr(1) |
| 2895 | .m(m) |
| 2896 | .n(n) |
| 2897 | .k(k) |
| 2898 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2899 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2900 | } |
| 2901 | } |
| 2902 | } |
| 2903 | } |
| 2904 | |
| 2905 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_gt_8) { |
| 2906 | TEST_REQUIRES_ARM_NEON; |
| 2907 | for (size_t k = 9; k < 16; k++) { |
| 2908 | GemmMicrokernelTester() |
| 2909 | .mr(2) |
| 2910 | .nr(8) |
| 2911 | .kr(2) |
| 2912 | .sr(1) |
| 2913 | .m(2) |
| 2914 | .n(8) |
| 2915 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2916 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2917 | } |
| 2918 | } |
| 2919 | |
| 2920 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_gt_8_strided_a) { |
| 2921 | TEST_REQUIRES_ARM_NEON; |
| 2922 | for (size_t k = 9; k < 16; k++) { |
| 2923 | GemmMicrokernelTester() |
| 2924 | .mr(2) |
| 2925 | .nr(8) |
| 2926 | .kr(2) |
| 2927 | .sr(1) |
| 2928 | .m(2) |
| 2929 | .n(8) |
| 2930 | .k(k) |
| 2931 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2932 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2933 | } |
| 2934 | } |
| 2935 | |
| 2936 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_gt_8_subtile) { |
| 2937 | TEST_REQUIRES_ARM_NEON; |
| 2938 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2939 | for (uint32_t n = 1; n <= 8; n++) { |
| 2940 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2941 | GemmMicrokernelTester() |
| 2942 | .mr(2) |
| 2943 | .nr(8) |
| 2944 | .kr(2) |
| 2945 | .sr(1) |
| 2946 | .m(m) |
| 2947 | .n(n) |
| 2948 | .k(k) |
| 2949 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2950 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2951 | } |
| 2952 | } |
| 2953 | } |
| 2954 | } |
| 2955 | |
| 2956 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_div_8) { |
| 2957 | TEST_REQUIRES_ARM_NEON; |
| 2958 | for (size_t k = 16; k <= 80; k += 8) { |
| 2959 | GemmMicrokernelTester() |
| 2960 | .mr(2) |
| 2961 | .nr(8) |
| 2962 | .kr(2) |
| 2963 | .sr(1) |
| 2964 | .m(2) |
| 2965 | .n(8) |
| 2966 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2967 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2968 | } |
| 2969 | } |
| 2970 | |
| 2971 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_div_8_strided_a) { |
| 2972 | TEST_REQUIRES_ARM_NEON; |
| 2973 | for (size_t k = 16; k <= 80; k += 8) { |
| 2974 | GemmMicrokernelTester() |
| 2975 | .mr(2) |
| 2976 | .nr(8) |
| 2977 | .kr(2) |
| 2978 | .sr(1) |
| 2979 | .m(2) |
| 2980 | .n(8) |
| 2981 | .k(k) |
| 2982 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 2983 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2984 | } |
| 2985 | } |
| 2986 | |
| 2987 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_div_8_subtile) { |
| 2988 | TEST_REQUIRES_ARM_NEON; |
| 2989 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 2990 | for (uint32_t n = 1; n <= 8; n++) { |
| 2991 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 2992 | GemmMicrokernelTester() |
| 2993 | .mr(2) |
| 2994 | .nr(8) |
| 2995 | .kr(2) |
| 2996 | .sr(1) |
| 2997 | .m(m) |
| 2998 | .n(n) |
| 2999 | .k(k) |
| 3000 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3001 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3002 | } |
| 3003 | } |
| 3004 | } |
| 3005 | } |
| 3006 | |
| 3007 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8) { |
| 3008 | TEST_REQUIRES_ARM_NEON; |
| 3009 | for (uint32_t n = 9; n < 16; n++) { |
| 3010 | for (size_t k = 1; k <= 40; k += 9) { |
| 3011 | GemmMicrokernelTester() |
| 3012 | .mr(2) |
| 3013 | .nr(8) |
| 3014 | .kr(2) |
| 3015 | .sr(1) |
| 3016 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3017 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3018 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3019 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3020 | } |
| 3021 | } |
| 3022 | } |
| 3023 | |
| 3024 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8_strided_cn) { |
| 3025 | TEST_REQUIRES_ARM_NEON; |
| 3026 | for (uint32_t n = 9; n < 16; n++) { |
| 3027 | for (size_t k = 1; k <= 40; k += 9) { |
| 3028 | GemmMicrokernelTester() |
| 3029 | .mr(2) |
| 3030 | .nr(8) |
| 3031 | .kr(2) |
| 3032 | .sr(1) |
| 3033 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3034 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3035 | .k(k) |
| 3036 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3037 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3038 | } |
| 3039 | } |
| 3040 | } |
| 3041 | |
| 3042 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8_strided_a) { |
| 3043 | TEST_REQUIRES_ARM_NEON; |
| 3044 | for (uint32_t n = 9; n < 16; n++) { |
| 3045 | for (size_t k = 1; k <= 40; k += 9) { |
| 3046 | GemmMicrokernelTester() |
| 3047 | .mr(2) |
| 3048 | .nr(8) |
| 3049 | .kr(2) |
| 3050 | .sr(1) |
| 3051 | .m(2) |
| 3052 | .n(n) |
| 3053 | .k(k) |
| 3054 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3055 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3056 | } |
| 3057 | } |
| 3058 | } |
| 3059 | |
| 3060 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8_subtile) { |
| 3061 | TEST_REQUIRES_ARM_NEON; |
| 3062 | for (uint32_t n = 9; n < 16; n++) { |
| 3063 | for (size_t k = 1; k <= 40; k += 9) { |
| 3064 | for (uint32_t m = 1; m <= 2; m++) { |
| 3065 | GemmMicrokernelTester() |
| 3066 | .mr(2) |
| 3067 | .nr(8) |
| 3068 | .kr(2) |
| 3069 | .sr(1) |
| 3070 | .m(m) |
| 3071 | .n(n) |
| 3072 | .k(k) |
| 3073 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3074 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3075 | } |
| 3076 | } |
| 3077 | } |
| 3078 | } |
| 3079 | |
| 3080 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8) { |
| 3081 | TEST_REQUIRES_ARM_NEON; |
| 3082 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3083 | for (size_t k = 1; k <= 40; k += 9) { |
| 3084 | GemmMicrokernelTester() |
| 3085 | .mr(2) |
| 3086 | .nr(8) |
| 3087 | .kr(2) |
| 3088 | .sr(1) |
| 3089 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3090 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3091 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3092 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3093 | } |
| 3094 | } |
| 3095 | } |
| 3096 | |
| 3097 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8_strided_cn) { |
| 3098 | TEST_REQUIRES_ARM_NEON; |
| 3099 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3100 | for (size_t k = 1; k <= 40; k += 9) { |
| 3101 | GemmMicrokernelTester() |
| 3102 | .mr(2) |
| 3103 | .nr(8) |
| 3104 | .kr(2) |
| 3105 | .sr(1) |
| 3106 | .m(2) |
| 3107 | .n(n) |
| 3108 | .k(k) |
| 3109 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3110 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3111 | } |
| 3112 | } |
| 3113 | } |
| 3114 | |
| 3115 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8_strided_a) { |
| 3116 | TEST_REQUIRES_ARM_NEON; |
| 3117 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3118 | for (size_t k = 1; k <= 40; k += 9) { |
| 3119 | GemmMicrokernelTester() |
| 3120 | .mr(2) |
| 3121 | .nr(8) |
| 3122 | .kr(2) |
| 3123 | .sr(1) |
| 3124 | .m(2) |
| 3125 | .n(n) |
| 3126 | .k(k) |
| 3127 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3128 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3129 | } |
| 3130 | } |
| 3131 | } |
| 3132 | |
| 3133 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8_subtile) { |
| 3134 | TEST_REQUIRES_ARM_NEON; |
| 3135 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3136 | for (size_t k = 1; k <= 40; k += 9) { |
| 3137 | for (uint32_t m = 1; m <= 2; m++) { |
| 3138 | GemmMicrokernelTester() |
| 3139 | .mr(2) |
| 3140 | .nr(8) |
| 3141 | .kr(2) |
| 3142 | .sr(1) |
| 3143 | .m(m) |
| 3144 | .n(n) |
| 3145 | .k(k) |
| 3146 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3147 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3148 | } |
| 3149 | } |
| 3150 | } |
| 3151 | } |
| 3152 | |
| 3153 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, strided_cm_subtile) { |
| 3154 | TEST_REQUIRES_ARM_NEON; |
| 3155 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3156 | for (uint32_t n = 1; n <= 8; n++) { |
| 3157 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3158 | GemmMicrokernelTester() |
| 3159 | .mr(2) |
| 3160 | .nr(8) |
| 3161 | .kr(2) |
| 3162 | .sr(1) |
| 3163 | .m(m) |
| 3164 | .n(n) |
| 3165 | .k(k) |
| 3166 | .cm_stride(11) |
| 3167 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3168 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3169 | } |
| 3170 | } |
| 3171 | } |
| 3172 | } |
| 3173 | |
| 3174 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, qmin) { |
| 3175 | TEST_REQUIRES_ARM_NEON; |
| 3176 | GemmMicrokernelTester() |
| 3177 | .mr(2) |
| 3178 | .nr(8) |
| 3179 | .kr(2) |
| 3180 | .sr(1) |
| 3181 | .m(2) |
| 3182 | .n(8) |
| 3183 | .k(8) |
| 3184 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3185 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3186 | } |
| 3187 | |
| 3188 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, qmax) { |
| 3189 | TEST_REQUIRES_ARM_NEON; |
| 3190 | GemmMicrokernelTester() |
| 3191 | .mr(2) |
| 3192 | .nr(8) |
| 3193 | .kr(2) |
| 3194 | .sr(1) |
| 3195 | .m(2) |
| 3196 | .n(8) |
| 3197 | .k(8) |
| 3198 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3199 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3200 | } |
| 3201 | |
| 3202 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, strided_cm) { |
| 3203 | TEST_REQUIRES_ARM_NEON; |
| 3204 | GemmMicrokernelTester() |
| 3205 | .mr(2) |
| 3206 | .nr(8) |
| 3207 | .kr(2) |
| 3208 | .sr(1) |
| 3209 | .m(2) |
| 3210 | .n(8) |
| 3211 | .k(8) |
| 3212 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3213 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3214 | } |
| 3215 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 3216 | |
| 3217 | |
| 3218 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 3219 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_eq_8) { |
| 3220 | TEST_REQUIRES_ARM_NEON; |
| 3221 | GemmMicrokernelTester() |
| 3222 | .mr(4) |
| 3223 | .nr(8) |
| 3224 | .kr(2) |
| 3225 | .sr(1) |
| 3226 | .m(4) |
| 3227 | .n(8) |
| 3228 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3229 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3230 | } |
| 3231 | |
| 3232 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, strided_cn) { |
| 3233 | TEST_REQUIRES_ARM_NEON; |
| 3234 | GemmMicrokernelTester() |
| 3235 | .mr(4) |
| 3236 | .nr(8) |
| 3237 | .kr(2) |
| 3238 | .sr(1) |
| 3239 | .m(4) |
| 3240 | .n(8) |
| 3241 | .k(8) |
| 3242 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3243 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3244 | } |
| 3245 | |
| 3246 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_eq_8_strided_a) { |
| 3247 | TEST_REQUIRES_ARM_NEON; |
| 3248 | GemmMicrokernelTester() |
| 3249 | .mr(4) |
| 3250 | .nr(8) |
| 3251 | .kr(2) |
| 3252 | .sr(1) |
| 3253 | .m(4) |
| 3254 | .n(8) |
| 3255 | .k(8) |
| 3256 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3257 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3258 | } |
| 3259 | |
| 3260 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_eq_8_subtile) { |
| 3261 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3262 | for (uint32_t n = 1; n <= 8; n++) { |
| 3263 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3264 | GemmMicrokernelTester() |
| 3265 | .mr(4) |
| 3266 | .nr(8) |
| 3267 | .kr(2) |
| 3268 | .sr(1) |
| 3269 | .m(m) |
| 3270 | .n(n) |
| 3271 | .k(8) |
| 3272 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3273 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3274 | } |
| 3275 | } |
| 3276 | } |
| 3277 | |
| 3278 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_eq_8_subtile_m) { |
| 3279 | TEST_REQUIRES_ARM_NEON; |
| 3280 | for (uint32_t m = 1; m <= 4; m++) { |
| 3281 | GemmMicrokernelTester() |
| 3282 | .mr(4) |
| 3283 | .nr(8) |
| 3284 | .kr(2) |
| 3285 | .sr(1) |
| 3286 | .m(m) |
| 3287 | .n(8) |
| 3288 | .k(8) |
| 3289 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3290 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3291 | } |
| 3292 | } |
| 3293 | |
| 3294 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_eq_8_subtile_n) { |
| 3295 | TEST_REQUIRES_ARM_NEON; |
| 3296 | for (uint32_t n = 1; n <= 8; n++) { |
| 3297 | GemmMicrokernelTester() |
| 3298 | .mr(4) |
| 3299 | .nr(8) |
| 3300 | .kr(2) |
| 3301 | .sr(1) |
| 3302 | .m(4) |
| 3303 | .n(n) |
| 3304 | .k(8) |
| 3305 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3306 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3307 | } |
| 3308 | } |
| 3309 | |
| 3310 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_lt_8) { |
| 3311 | TEST_REQUIRES_ARM_NEON; |
| 3312 | for (size_t k = 1; k < 8; k++) { |
| 3313 | GemmMicrokernelTester() |
| 3314 | .mr(4) |
| 3315 | .nr(8) |
| 3316 | .kr(2) |
| 3317 | .sr(1) |
| 3318 | .m(4) |
| 3319 | .n(8) |
| 3320 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3321 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3322 | } |
| 3323 | } |
| 3324 | |
| 3325 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_lt_8_strided_a) { |
| 3326 | TEST_REQUIRES_ARM_NEON; |
| 3327 | for (size_t k = 1; k < 8; k++) { |
| 3328 | GemmMicrokernelTester() |
| 3329 | .mr(4) |
| 3330 | .nr(8) |
| 3331 | .kr(2) |
| 3332 | .sr(1) |
| 3333 | .m(4) |
| 3334 | .n(8) |
| 3335 | .k(k) |
| 3336 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3337 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3338 | } |
| 3339 | } |
| 3340 | |
| 3341 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_lt_8_subtile) { |
| 3342 | TEST_REQUIRES_ARM_NEON; |
| 3343 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3344 | for (uint32_t n = 1; n <= 8; n++) { |
| 3345 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3346 | GemmMicrokernelTester() |
| 3347 | .mr(4) |
| 3348 | .nr(8) |
| 3349 | .kr(2) |
| 3350 | .sr(1) |
| 3351 | .m(m) |
| 3352 | .n(n) |
| 3353 | .k(k) |
| 3354 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3355 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3356 | } |
| 3357 | } |
| 3358 | } |
| 3359 | } |
| 3360 | |
| 3361 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_gt_8) { |
| 3362 | TEST_REQUIRES_ARM_NEON; |
| 3363 | for (size_t k = 9; k < 16; k++) { |
| 3364 | GemmMicrokernelTester() |
| 3365 | .mr(4) |
| 3366 | .nr(8) |
| 3367 | .kr(2) |
| 3368 | .sr(1) |
| 3369 | .m(4) |
| 3370 | .n(8) |
| 3371 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3372 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3373 | } |
| 3374 | } |
| 3375 | |
| 3376 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_gt_8_strided_a) { |
| 3377 | TEST_REQUIRES_ARM_NEON; |
| 3378 | for (size_t k = 9; k < 16; k++) { |
| 3379 | GemmMicrokernelTester() |
| 3380 | .mr(4) |
| 3381 | .nr(8) |
| 3382 | .kr(2) |
| 3383 | .sr(1) |
| 3384 | .m(4) |
| 3385 | .n(8) |
| 3386 | .k(k) |
| 3387 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3388 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3389 | } |
| 3390 | } |
| 3391 | |
| 3392 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_gt_8_subtile) { |
| 3393 | TEST_REQUIRES_ARM_NEON; |
| 3394 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3395 | for (uint32_t n = 1; n <= 8; n++) { |
| 3396 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3397 | GemmMicrokernelTester() |
| 3398 | .mr(4) |
| 3399 | .nr(8) |
| 3400 | .kr(2) |
| 3401 | .sr(1) |
| 3402 | .m(m) |
| 3403 | .n(n) |
| 3404 | .k(k) |
| 3405 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3406 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3407 | } |
| 3408 | } |
| 3409 | } |
| 3410 | } |
| 3411 | |
| 3412 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_div_8) { |
| 3413 | TEST_REQUIRES_ARM_NEON; |
| 3414 | for (size_t k = 16; k <= 80; k += 8) { |
| 3415 | GemmMicrokernelTester() |
| 3416 | .mr(4) |
| 3417 | .nr(8) |
| 3418 | .kr(2) |
| 3419 | .sr(1) |
| 3420 | .m(4) |
| 3421 | .n(8) |
| 3422 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3423 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3424 | } |
| 3425 | } |
| 3426 | |
| 3427 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_div_8_strided_a) { |
| 3428 | TEST_REQUIRES_ARM_NEON; |
| 3429 | for (size_t k = 16; k <= 80; k += 8) { |
| 3430 | GemmMicrokernelTester() |
| 3431 | .mr(4) |
| 3432 | .nr(8) |
| 3433 | .kr(2) |
| 3434 | .sr(1) |
| 3435 | .m(4) |
| 3436 | .n(8) |
| 3437 | .k(k) |
| 3438 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3439 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3440 | } |
| 3441 | } |
| 3442 | |
| 3443 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, k_div_8_subtile) { |
| 3444 | TEST_REQUIRES_ARM_NEON; |
| 3445 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3446 | for (uint32_t n = 1; n <= 8; n++) { |
| 3447 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3448 | GemmMicrokernelTester() |
| 3449 | .mr(4) |
| 3450 | .nr(8) |
| 3451 | .kr(2) |
| 3452 | .sr(1) |
| 3453 | .m(m) |
| 3454 | .n(n) |
| 3455 | .k(k) |
| 3456 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3457 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3458 | } |
| 3459 | } |
| 3460 | } |
| 3461 | } |
| 3462 | |
| 3463 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_gt_8) { |
| 3464 | TEST_REQUIRES_ARM_NEON; |
| 3465 | for (uint32_t n = 9; n < 16; n++) { |
| 3466 | for (size_t k = 1; k <= 40; k += 9) { |
| 3467 | GemmMicrokernelTester() |
| 3468 | .mr(4) |
| 3469 | .nr(8) |
| 3470 | .kr(2) |
| 3471 | .sr(1) |
| 3472 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3473 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3474 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3475 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3476 | } |
| 3477 | } |
| 3478 | } |
| 3479 | |
| 3480 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_gt_8_strided_cn) { |
| 3481 | TEST_REQUIRES_ARM_NEON; |
| 3482 | for (uint32_t n = 9; n < 16; n++) { |
| 3483 | for (size_t k = 1; k <= 40; k += 9) { |
| 3484 | GemmMicrokernelTester() |
| 3485 | .mr(4) |
| 3486 | .nr(8) |
| 3487 | .kr(2) |
| 3488 | .sr(1) |
| 3489 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3490 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3491 | .k(k) |
| 3492 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3493 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3494 | } |
| 3495 | } |
| 3496 | } |
| 3497 | |
| 3498 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_gt_8_strided_a) { |
| 3499 | TEST_REQUIRES_ARM_NEON; |
| 3500 | for (uint32_t n = 9; n < 16; n++) { |
| 3501 | for (size_t k = 1; k <= 40; k += 9) { |
| 3502 | GemmMicrokernelTester() |
| 3503 | .mr(4) |
| 3504 | .nr(8) |
| 3505 | .kr(2) |
| 3506 | .sr(1) |
| 3507 | .m(4) |
| 3508 | .n(n) |
| 3509 | .k(k) |
| 3510 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3511 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3512 | } |
| 3513 | } |
| 3514 | } |
| 3515 | |
| 3516 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_gt_8_subtile) { |
| 3517 | TEST_REQUIRES_ARM_NEON; |
| 3518 | for (uint32_t n = 9; n < 16; n++) { |
| 3519 | for (size_t k = 1; k <= 40; k += 9) { |
| 3520 | for (uint32_t m = 1; m <= 4; m++) { |
| 3521 | GemmMicrokernelTester() |
| 3522 | .mr(4) |
| 3523 | .nr(8) |
| 3524 | .kr(2) |
| 3525 | .sr(1) |
| 3526 | .m(m) |
| 3527 | .n(n) |
| 3528 | .k(k) |
| 3529 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3530 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3531 | } |
| 3532 | } |
| 3533 | } |
| 3534 | } |
| 3535 | |
| 3536 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_div_8) { |
| 3537 | TEST_REQUIRES_ARM_NEON; |
| 3538 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3539 | for (size_t k = 1; k <= 40; k += 9) { |
| 3540 | GemmMicrokernelTester() |
| 3541 | .mr(4) |
| 3542 | .nr(8) |
| 3543 | .kr(2) |
| 3544 | .sr(1) |
| 3545 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3546 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3547 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3548 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3549 | } |
| 3550 | } |
| 3551 | } |
| 3552 | |
| 3553 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_div_8_strided_cn) { |
| 3554 | TEST_REQUIRES_ARM_NEON; |
| 3555 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3556 | for (size_t k = 1; k <= 40; k += 9) { |
| 3557 | GemmMicrokernelTester() |
| 3558 | .mr(4) |
| 3559 | .nr(8) |
| 3560 | .kr(2) |
| 3561 | .sr(1) |
| 3562 | .m(4) |
| 3563 | .n(n) |
| 3564 | .k(k) |
| 3565 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3566 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3567 | } |
| 3568 | } |
| 3569 | } |
| 3570 | |
| 3571 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_div_8_strided_a) { |
| 3572 | TEST_REQUIRES_ARM_NEON; |
| 3573 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3574 | for (size_t k = 1; k <= 40; k += 9) { |
| 3575 | GemmMicrokernelTester() |
| 3576 | .mr(4) |
| 3577 | .nr(8) |
| 3578 | .kr(2) |
| 3579 | .sr(1) |
| 3580 | .m(4) |
| 3581 | .n(n) |
| 3582 | .k(k) |
| 3583 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3584 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3585 | } |
| 3586 | } |
| 3587 | } |
| 3588 | |
| 3589 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, n_div_8_subtile) { |
| 3590 | TEST_REQUIRES_ARM_NEON; |
| 3591 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 3592 | for (size_t k = 1; k <= 40; k += 9) { |
| 3593 | for (uint32_t m = 1; m <= 4; m++) { |
| 3594 | GemmMicrokernelTester() |
| 3595 | .mr(4) |
| 3596 | .nr(8) |
| 3597 | .kr(2) |
| 3598 | .sr(1) |
| 3599 | .m(m) |
| 3600 | .n(n) |
| 3601 | .k(k) |
| 3602 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3603 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3604 | } |
| 3605 | } |
| 3606 | } |
| 3607 | } |
| 3608 | |
| 3609 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, strided_cm_subtile) { |
| 3610 | TEST_REQUIRES_ARM_NEON; |
| 3611 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3612 | for (uint32_t n = 1; n <= 8; n++) { |
| 3613 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3614 | GemmMicrokernelTester() |
| 3615 | .mr(4) |
| 3616 | .nr(8) |
| 3617 | .kr(2) |
| 3618 | .sr(1) |
| 3619 | .m(m) |
| 3620 | .n(n) |
| 3621 | .k(k) |
| 3622 | .cm_stride(11) |
| 3623 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3624 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3625 | } |
| 3626 | } |
| 3627 | } |
| 3628 | } |
| 3629 | |
| 3630 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, qmin) { |
| 3631 | TEST_REQUIRES_ARM_NEON; |
| 3632 | GemmMicrokernelTester() |
| 3633 | .mr(4) |
| 3634 | .nr(8) |
| 3635 | .kr(2) |
| 3636 | .sr(1) |
| 3637 | .m(4) |
| 3638 | .n(8) |
| 3639 | .k(8) |
| 3640 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3641 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3642 | } |
| 3643 | |
| 3644 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, qmax) { |
| 3645 | TEST_REQUIRES_ARM_NEON; |
| 3646 | GemmMicrokernelTester() |
| 3647 | .mr(4) |
| 3648 | .nr(8) |
| 3649 | .kr(2) |
| 3650 | .sr(1) |
| 3651 | .m(4) |
| 3652 | .n(8) |
| 3653 | .k(8) |
| 3654 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3655 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3656 | } |
| 3657 | |
| 3658 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD2R, strided_cm) { |
| 3659 | TEST_REQUIRES_ARM_NEON; |
| 3660 | GemmMicrokernelTester() |
| 3661 | .mr(4) |
| 3662 | .nr(8) |
| 3663 | .kr(2) |
| 3664 | .sr(1) |
| 3665 | .m(4) |
| 3666 | .n(8) |
| 3667 | .k(8) |
| 3668 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3669 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3670 | } |
| 3671 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 3672 | |
| 3673 | |
| 3674 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 3675 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8) { |
| 3676 | TEST_REQUIRES_ARM_NEON; |
| 3677 | GemmMicrokernelTester() |
| 3678 | .mr(3) |
| 3679 | .nr(16) |
| 3680 | .kr(2) |
| 3681 | .sr(1) |
| 3682 | .m(3) |
| 3683 | .n(16) |
| 3684 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3685 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3686 | } |
| 3687 | |
| 3688 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, strided_cn) { |
| 3689 | TEST_REQUIRES_ARM_NEON; |
| 3690 | GemmMicrokernelTester() |
| 3691 | .mr(3) |
| 3692 | .nr(16) |
| 3693 | .kr(2) |
| 3694 | .sr(1) |
| 3695 | .m(3) |
| 3696 | .n(16) |
| 3697 | .k(8) |
| 3698 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3699 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3700 | } |
| 3701 | |
| 3702 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8_strided_a) { |
| 3703 | TEST_REQUIRES_ARM_NEON; |
| 3704 | GemmMicrokernelTester() |
| 3705 | .mr(3) |
| 3706 | .nr(16) |
| 3707 | .kr(2) |
| 3708 | .sr(1) |
| 3709 | .m(3) |
| 3710 | .n(16) |
| 3711 | .k(8) |
| 3712 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3713 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3714 | } |
| 3715 | |
| 3716 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8_subtile) { |
| 3717 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3718 | for (uint32_t n = 1; n <= 16; n++) { |
| 3719 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3720 | GemmMicrokernelTester() |
| 3721 | .mr(3) |
| 3722 | .nr(16) |
| 3723 | .kr(2) |
| 3724 | .sr(1) |
| 3725 | .m(m) |
| 3726 | .n(n) |
| 3727 | .k(8) |
| 3728 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3729 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3730 | } |
| 3731 | } |
| 3732 | } |
| 3733 | |
| 3734 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8_subtile_m) { |
| 3735 | TEST_REQUIRES_ARM_NEON; |
| 3736 | for (uint32_t m = 1; m <= 3; m++) { |
| 3737 | GemmMicrokernelTester() |
| 3738 | .mr(3) |
| 3739 | .nr(16) |
| 3740 | .kr(2) |
| 3741 | .sr(1) |
| 3742 | .m(m) |
| 3743 | .n(16) |
| 3744 | .k(8) |
| 3745 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3746 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3747 | } |
| 3748 | } |
| 3749 | |
| 3750 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_eq_8_subtile_n) { |
| 3751 | TEST_REQUIRES_ARM_NEON; |
| 3752 | for (uint32_t n = 1; n <= 16; n++) { |
| 3753 | GemmMicrokernelTester() |
| 3754 | .mr(3) |
| 3755 | .nr(16) |
| 3756 | .kr(2) |
| 3757 | .sr(1) |
| 3758 | .m(3) |
| 3759 | .n(n) |
| 3760 | .k(8) |
| 3761 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3762 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3763 | } |
| 3764 | } |
| 3765 | |
| 3766 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_lt_8) { |
| 3767 | TEST_REQUIRES_ARM_NEON; |
| 3768 | for (size_t k = 1; k < 8; k++) { |
| 3769 | GemmMicrokernelTester() |
| 3770 | .mr(3) |
| 3771 | .nr(16) |
| 3772 | .kr(2) |
| 3773 | .sr(1) |
| 3774 | .m(3) |
| 3775 | .n(16) |
| 3776 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3777 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3778 | } |
| 3779 | } |
| 3780 | |
| 3781 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_lt_8_strided_a) { |
| 3782 | TEST_REQUIRES_ARM_NEON; |
| 3783 | for (size_t k = 1; k < 8; k++) { |
| 3784 | GemmMicrokernelTester() |
| 3785 | .mr(3) |
| 3786 | .nr(16) |
| 3787 | .kr(2) |
| 3788 | .sr(1) |
| 3789 | .m(3) |
| 3790 | .n(16) |
| 3791 | .k(k) |
| 3792 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3793 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3794 | } |
| 3795 | } |
| 3796 | |
| 3797 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_lt_8_subtile) { |
| 3798 | TEST_REQUIRES_ARM_NEON; |
| 3799 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3800 | for (uint32_t n = 1; n <= 16; n++) { |
| 3801 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3802 | GemmMicrokernelTester() |
| 3803 | .mr(3) |
| 3804 | .nr(16) |
| 3805 | .kr(2) |
| 3806 | .sr(1) |
| 3807 | .m(m) |
| 3808 | .n(n) |
| 3809 | .k(k) |
| 3810 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3811 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3812 | } |
| 3813 | } |
| 3814 | } |
| 3815 | } |
| 3816 | |
| 3817 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_gt_8) { |
| 3818 | TEST_REQUIRES_ARM_NEON; |
| 3819 | for (size_t k = 9; k < 16; k++) { |
| 3820 | GemmMicrokernelTester() |
| 3821 | .mr(3) |
| 3822 | .nr(16) |
| 3823 | .kr(2) |
| 3824 | .sr(1) |
| 3825 | .m(3) |
| 3826 | .n(16) |
| 3827 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3828 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3829 | } |
| 3830 | } |
| 3831 | |
| 3832 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_gt_8_strided_a) { |
| 3833 | TEST_REQUIRES_ARM_NEON; |
| 3834 | for (size_t k = 9; k < 16; k++) { |
| 3835 | GemmMicrokernelTester() |
| 3836 | .mr(3) |
| 3837 | .nr(16) |
| 3838 | .kr(2) |
| 3839 | .sr(1) |
| 3840 | .m(3) |
| 3841 | .n(16) |
| 3842 | .k(k) |
| 3843 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3844 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3845 | } |
| 3846 | } |
| 3847 | |
| 3848 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_gt_8_subtile) { |
| 3849 | TEST_REQUIRES_ARM_NEON; |
| 3850 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3851 | for (uint32_t n = 1; n <= 16; n++) { |
| 3852 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3853 | GemmMicrokernelTester() |
| 3854 | .mr(3) |
| 3855 | .nr(16) |
| 3856 | .kr(2) |
| 3857 | .sr(1) |
| 3858 | .m(m) |
| 3859 | .n(n) |
| 3860 | .k(k) |
| 3861 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3862 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3863 | } |
| 3864 | } |
| 3865 | } |
| 3866 | } |
| 3867 | |
| 3868 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_div_8) { |
| 3869 | TEST_REQUIRES_ARM_NEON; |
| 3870 | for (size_t k = 16; k <= 80; k += 8) { |
| 3871 | GemmMicrokernelTester() |
| 3872 | .mr(3) |
| 3873 | .nr(16) |
| 3874 | .kr(2) |
| 3875 | .sr(1) |
| 3876 | .m(3) |
| 3877 | .n(16) |
| 3878 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3879 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3880 | } |
| 3881 | } |
| 3882 | |
| 3883 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_div_8_strided_a) { |
| 3884 | TEST_REQUIRES_ARM_NEON; |
| 3885 | for (size_t k = 16; k <= 80; k += 8) { |
| 3886 | GemmMicrokernelTester() |
| 3887 | .mr(3) |
| 3888 | .nr(16) |
| 3889 | .kr(2) |
| 3890 | .sr(1) |
| 3891 | .m(3) |
| 3892 | .n(16) |
| 3893 | .k(k) |
| 3894 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3895 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3896 | } |
| 3897 | } |
| 3898 | |
| 3899 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, k_div_8_subtile) { |
| 3900 | TEST_REQUIRES_ARM_NEON; |
| 3901 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 3902 | for (uint32_t n = 1; n <= 16; n++) { |
| 3903 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3904 | GemmMicrokernelTester() |
| 3905 | .mr(3) |
| 3906 | .nr(16) |
| 3907 | .kr(2) |
| 3908 | .sr(1) |
| 3909 | .m(m) |
| 3910 | .n(n) |
| 3911 | .k(k) |
| 3912 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3913 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3914 | } |
| 3915 | } |
| 3916 | } |
| 3917 | } |
| 3918 | |
| 3919 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16) { |
| 3920 | TEST_REQUIRES_ARM_NEON; |
| 3921 | for (uint32_t n = 17; n < 32; n++) { |
| 3922 | for (size_t k = 1; k <= 40; k += 9) { |
| 3923 | GemmMicrokernelTester() |
| 3924 | .mr(3) |
| 3925 | .nr(16) |
| 3926 | .kr(2) |
| 3927 | .sr(1) |
| 3928 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3929 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3930 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3931 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3932 | } |
| 3933 | } |
| 3934 | } |
| 3935 | |
| 3936 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16_strided_cn) { |
| 3937 | TEST_REQUIRES_ARM_NEON; |
| 3938 | for (uint32_t n = 17; n < 32; n++) { |
| 3939 | for (size_t k = 1; k <= 40; k += 9) { |
| 3940 | GemmMicrokernelTester() |
| 3941 | .mr(3) |
| 3942 | .nr(16) |
| 3943 | .kr(2) |
| 3944 | .sr(1) |
| 3945 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 3946 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3947 | .k(k) |
| 3948 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3949 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3950 | } |
| 3951 | } |
| 3952 | } |
| 3953 | |
| 3954 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16_strided_a) { |
| 3955 | TEST_REQUIRES_ARM_NEON; |
| 3956 | for (uint32_t n = 17; n < 32; n++) { |
| 3957 | for (size_t k = 1; k <= 40; k += 9) { |
| 3958 | GemmMicrokernelTester() |
| 3959 | .mr(3) |
| 3960 | .nr(16) |
| 3961 | .kr(2) |
| 3962 | .sr(1) |
| 3963 | .m(3) |
| 3964 | .n(n) |
| 3965 | .k(k) |
| 3966 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3967 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3968 | } |
| 3969 | } |
| 3970 | } |
| 3971 | |
| 3972 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_gt_16_subtile) { |
| 3973 | TEST_REQUIRES_ARM_NEON; |
| 3974 | for (uint32_t n = 17; n < 32; n++) { |
| 3975 | for (size_t k = 1; k <= 40; k += 9) { |
| 3976 | for (uint32_t m = 1; m <= 3; m++) { |
| 3977 | GemmMicrokernelTester() |
| 3978 | .mr(3) |
| 3979 | .nr(16) |
| 3980 | .kr(2) |
| 3981 | .sr(1) |
| 3982 | .m(m) |
| 3983 | .n(n) |
| 3984 | .k(k) |
| 3985 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 3986 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 3987 | } |
| 3988 | } |
| 3989 | } |
| 3990 | } |
| 3991 | |
| 3992 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16) { |
| 3993 | TEST_REQUIRES_ARM_NEON; |
| 3994 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 3995 | for (size_t k = 1; k <= 40; k += 9) { |
| 3996 | GemmMicrokernelTester() |
| 3997 | .mr(3) |
| 3998 | .nr(16) |
| 3999 | .kr(2) |
| 4000 | .sr(1) |
| 4001 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4002 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4003 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4004 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4005 | } |
| 4006 | } |
| 4007 | } |
| 4008 | |
| 4009 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16_strided_cn) { |
| 4010 | TEST_REQUIRES_ARM_NEON; |
| 4011 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4012 | for (size_t k = 1; k <= 40; k += 9) { |
| 4013 | GemmMicrokernelTester() |
| 4014 | .mr(3) |
| 4015 | .nr(16) |
| 4016 | .kr(2) |
| 4017 | .sr(1) |
| 4018 | .m(3) |
| 4019 | .n(n) |
| 4020 | .k(k) |
| 4021 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4022 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4023 | } |
| 4024 | } |
| 4025 | } |
| 4026 | |
| 4027 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16_strided_a) { |
| 4028 | TEST_REQUIRES_ARM_NEON; |
| 4029 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4030 | for (size_t k = 1; k <= 40; k += 9) { |
| 4031 | GemmMicrokernelTester() |
| 4032 | .mr(3) |
| 4033 | .nr(16) |
| 4034 | .kr(2) |
| 4035 | .sr(1) |
| 4036 | .m(3) |
| 4037 | .n(n) |
| 4038 | .k(k) |
| 4039 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4040 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4041 | } |
| 4042 | } |
| 4043 | } |
| 4044 | |
| 4045 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, n_div_16_subtile) { |
| 4046 | TEST_REQUIRES_ARM_NEON; |
| 4047 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 4048 | for (size_t k = 1; k <= 40; k += 9) { |
| 4049 | for (uint32_t m = 1; m <= 3; m++) { |
| 4050 | GemmMicrokernelTester() |
| 4051 | .mr(3) |
| 4052 | .nr(16) |
| 4053 | .kr(2) |
| 4054 | .sr(1) |
| 4055 | .m(m) |
| 4056 | .n(n) |
| 4057 | .k(k) |
| 4058 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4059 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4060 | } |
| 4061 | } |
| 4062 | } |
| 4063 | } |
| 4064 | |
| 4065 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, strided_cm_subtile) { |
| 4066 | TEST_REQUIRES_ARM_NEON; |
| 4067 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4068 | for (uint32_t n = 1; n <= 16; n++) { |
| 4069 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4070 | GemmMicrokernelTester() |
| 4071 | .mr(3) |
| 4072 | .nr(16) |
| 4073 | .kr(2) |
| 4074 | .sr(1) |
| 4075 | .m(m) |
| 4076 | .n(n) |
| 4077 | .k(k) |
| 4078 | .cm_stride(19) |
| 4079 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4080 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4081 | } |
| 4082 | } |
| 4083 | } |
| 4084 | } |
| 4085 | |
| 4086 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, qmin) { |
| 4087 | TEST_REQUIRES_ARM_NEON; |
| 4088 | GemmMicrokernelTester() |
| 4089 | .mr(3) |
| 4090 | .nr(16) |
| 4091 | .kr(2) |
| 4092 | .sr(1) |
| 4093 | .m(3) |
| 4094 | .n(16) |
| 4095 | .k(8) |
| 4096 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4097 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4098 | } |
| 4099 | |
| 4100 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, qmax) { |
| 4101 | TEST_REQUIRES_ARM_NEON; |
| 4102 | GemmMicrokernelTester() |
| 4103 | .mr(3) |
| 4104 | .nr(16) |
| 4105 | .kr(2) |
| 4106 | .sr(1) |
| 4107 | .m(3) |
| 4108 | .n(16) |
| 4109 | .k(8) |
| 4110 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4111 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4112 | } |
| 4113 | |
| 4114 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_LD2R, strided_cm) { |
| 4115 | TEST_REQUIRES_ARM_NEON; |
| 4116 | GemmMicrokernelTester() |
| 4117 | .mr(3) |
| 4118 | .nr(16) |
| 4119 | .kr(2) |
| 4120 | .sr(1) |
| 4121 | .m(3) |
| 4122 | .n(16) |
| 4123 | .k(8) |
| 4124 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4125 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4126 | } |
| 4127 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4128 | |
| 4129 | |
| 4130 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4131 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_eq_16) { |
| 4132 | TEST_REQUIRES_ARM_NEON; |
| 4133 | GemmMicrokernelTester() |
| 4134 | .mr(1) |
| 4135 | .nr(8) |
| 4136 | .kr(2) |
| 4137 | .sr(1) |
| 4138 | .m(1) |
| 4139 | .n(8) |
| 4140 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4141 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4142 | } |
| 4143 | |
| 4144 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, strided_cn) { |
| 4145 | TEST_REQUIRES_ARM_NEON; |
| 4146 | GemmMicrokernelTester() |
| 4147 | .mr(1) |
| 4148 | .nr(8) |
| 4149 | .kr(2) |
| 4150 | .sr(1) |
| 4151 | .m(1) |
| 4152 | .n(8) |
| 4153 | .k(16) |
| 4154 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4155 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4156 | } |
| 4157 | |
| 4158 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_eq_16_strided_a) { |
| 4159 | TEST_REQUIRES_ARM_NEON; |
| 4160 | GemmMicrokernelTester() |
| 4161 | .mr(1) |
| 4162 | .nr(8) |
| 4163 | .kr(2) |
| 4164 | .sr(1) |
| 4165 | .m(1) |
| 4166 | .n(8) |
| 4167 | .k(16) |
| 4168 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4169 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4170 | } |
| 4171 | |
| 4172 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_eq_16_subtile) { |
| 4173 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4174 | for (uint32_t n = 1; n <= 8; n++) { |
| 4175 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4176 | GemmMicrokernelTester() |
| 4177 | .mr(1) |
| 4178 | .nr(8) |
| 4179 | .kr(2) |
| 4180 | .sr(1) |
| 4181 | .m(m) |
| 4182 | .n(n) |
| 4183 | .k(16) |
| 4184 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4185 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4186 | } |
| 4187 | } |
| 4188 | } |
| 4189 | |
| 4190 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_eq_16_subtile_m) { |
| 4191 | TEST_REQUIRES_ARM_NEON; |
| 4192 | for (uint32_t m = 1; m <= 1; m++) { |
| 4193 | GemmMicrokernelTester() |
| 4194 | .mr(1) |
| 4195 | .nr(8) |
| 4196 | .kr(2) |
| 4197 | .sr(1) |
| 4198 | .m(m) |
| 4199 | .n(8) |
| 4200 | .k(16) |
| 4201 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4202 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4203 | } |
| 4204 | } |
| 4205 | |
| 4206 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_eq_16_subtile_n) { |
| 4207 | TEST_REQUIRES_ARM_NEON; |
| 4208 | for (uint32_t n = 1; n <= 8; n++) { |
| 4209 | GemmMicrokernelTester() |
| 4210 | .mr(1) |
| 4211 | .nr(8) |
| 4212 | .kr(2) |
| 4213 | .sr(1) |
| 4214 | .m(1) |
| 4215 | .n(n) |
| 4216 | .k(16) |
| 4217 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4218 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4219 | } |
| 4220 | } |
| 4221 | |
| 4222 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_lt_16) { |
| 4223 | TEST_REQUIRES_ARM_NEON; |
| 4224 | for (size_t k = 1; k < 16; k++) { |
| 4225 | GemmMicrokernelTester() |
| 4226 | .mr(1) |
| 4227 | .nr(8) |
| 4228 | .kr(2) |
| 4229 | .sr(1) |
| 4230 | .m(1) |
| 4231 | .n(8) |
| 4232 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4233 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4234 | } |
| 4235 | } |
| 4236 | |
| 4237 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_lt_16_strided_a) { |
| 4238 | TEST_REQUIRES_ARM_NEON; |
| 4239 | for (size_t k = 1; k < 16; k++) { |
| 4240 | GemmMicrokernelTester() |
| 4241 | .mr(1) |
| 4242 | .nr(8) |
| 4243 | .kr(2) |
| 4244 | .sr(1) |
| 4245 | .m(1) |
| 4246 | .n(8) |
| 4247 | .k(k) |
| 4248 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4249 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4250 | } |
| 4251 | } |
| 4252 | |
| 4253 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_lt_16_subtile) { |
| 4254 | TEST_REQUIRES_ARM_NEON; |
| 4255 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4256 | for (uint32_t n = 1; n <= 8; n++) { |
| 4257 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4258 | GemmMicrokernelTester() |
| 4259 | .mr(1) |
| 4260 | .nr(8) |
| 4261 | .kr(2) |
| 4262 | .sr(1) |
| 4263 | .m(m) |
| 4264 | .n(n) |
| 4265 | .k(k) |
| 4266 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4267 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4268 | } |
| 4269 | } |
| 4270 | } |
| 4271 | } |
| 4272 | |
| 4273 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_gt_16) { |
| 4274 | TEST_REQUIRES_ARM_NEON; |
| 4275 | for (size_t k = 17; k < 32; k++) { |
| 4276 | GemmMicrokernelTester() |
| 4277 | .mr(1) |
| 4278 | .nr(8) |
| 4279 | .kr(2) |
| 4280 | .sr(1) |
| 4281 | .m(1) |
| 4282 | .n(8) |
| 4283 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4284 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4285 | } |
| 4286 | } |
| 4287 | |
| 4288 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_gt_16_strided_a) { |
| 4289 | TEST_REQUIRES_ARM_NEON; |
| 4290 | for (size_t k = 17; k < 32; k++) { |
| 4291 | GemmMicrokernelTester() |
| 4292 | .mr(1) |
| 4293 | .nr(8) |
| 4294 | .kr(2) |
| 4295 | .sr(1) |
| 4296 | .m(1) |
| 4297 | .n(8) |
| 4298 | .k(k) |
| 4299 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4300 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4301 | } |
| 4302 | } |
| 4303 | |
| 4304 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_gt_16_subtile) { |
| 4305 | TEST_REQUIRES_ARM_NEON; |
| 4306 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4307 | for (uint32_t n = 1; n <= 8; n++) { |
| 4308 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4309 | GemmMicrokernelTester() |
| 4310 | .mr(1) |
| 4311 | .nr(8) |
| 4312 | .kr(2) |
| 4313 | .sr(1) |
| 4314 | .m(m) |
| 4315 | .n(n) |
| 4316 | .k(k) |
| 4317 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4318 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4319 | } |
| 4320 | } |
| 4321 | } |
| 4322 | } |
| 4323 | |
| 4324 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_div_16) { |
| 4325 | TEST_REQUIRES_ARM_NEON; |
| 4326 | for (size_t k = 32; k <= 160; k += 16) { |
| 4327 | GemmMicrokernelTester() |
| 4328 | .mr(1) |
| 4329 | .nr(8) |
| 4330 | .kr(2) |
| 4331 | .sr(1) |
| 4332 | .m(1) |
| 4333 | .n(8) |
| 4334 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4335 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4336 | } |
| 4337 | } |
| 4338 | |
| 4339 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_div_16_strided_a) { |
| 4340 | TEST_REQUIRES_ARM_NEON; |
| 4341 | for (size_t k = 32; k <= 160; k += 16) { |
| 4342 | GemmMicrokernelTester() |
| 4343 | .mr(1) |
| 4344 | .nr(8) |
| 4345 | .kr(2) |
| 4346 | .sr(1) |
| 4347 | .m(1) |
| 4348 | .n(8) |
| 4349 | .k(k) |
| 4350 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4351 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4352 | } |
| 4353 | } |
| 4354 | |
| 4355 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, k_div_16_subtile) { |
| 4356 | TEST_REQUIRES_ARM_NEON; |
| 4357 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4358 | for (uint32_t n = 1; n <= 8; n++) { |
| 4359 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4360 | GemmMicrokernelTester() |
| 4361 | .mr(1) |
| 4362 | .nr(8) |
| 4363 | .kr(2) |
| 4364 | .sr(1) |
| 4365 | .m(m) |
| 4366 | .n(n) |
| 4367 | .k(k) |
| 4368 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4369 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4370 | } |
| 4371 | } |
| 4372 | } |
| 4373 | } |
| 4374 | |
| 4375 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_gt_8) { |
| 4376 | TEST_REQUIRES_ARM_NEON; |
| 4377 | for (uint32_t n = 9; n < 16; n++) { |
| 4378 | for (size_t k = 1; k <= 80; k += 17) { |
| 4379 | GemmMicrokernelTester() |
| 4380 | .mr(1) |
| 4381 | .nr(8) |
| 4382 | .kr(2) |
| 4383 | .sr(1) |
| 4384 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4385 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4386 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4387 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4388 | } |
| 4389 | } |
| 4390 | } |
| 4391 | |
| 4392 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_gt_8_strided_cn) { |
| 4393 | TEST_REQUIRES_ARM_NEON; |
| 4394 | for (uint32_t n = 9; n < 16; n++) { |
| 4395 | for (size_t k = 1; k <= 80; k += 17) { |
| 4396 | GemmMicrokernelTester() |
| 4397 | .mr(1) |
| 4398 | .nr(8) |
| 4399 | .kr(2) |
| 4400 | .sr(1) |
| 4401 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4402 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4403 | .k(k) |
| 4404 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4405 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4406 | } |
| 4407 | } |
| 4408 | } |
| 4409 | |
| 4410 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_gt_8_strided_a) { |
| 4411 | TEST_REQUIRES_ARM_NEON; |
| 4412 | for (uint32_t n = 9; n < 16; n++) { |
| 4413 | for (size_t k = 1; k <= 80; k += 17) { |
| 4414 | GemmMicrokernelTester() |
| 4415 | .mr(1) |
| 4416 | .nr(8) |
| 4417 | .kr(2) |
| 4418 | .sr(1) |
| 4419 | .m(1) |
| 4420 | .n(n) |
| 4421 | .k(k) |
| 4422 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4423 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4424 | } |
| 4425 | } |
| 4426 | } |
| 4427 | |
| 4428 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_gt_8_subtile) { |
| 4429 | TEST_REQUIRES_ARM_NEON; |
| 4430 | for (uint32_t n = 9; n < 16; n++) { |
| 4431 | for (size_t k = 1; k <= 80; k += 17) { |
| 4432 | for (uint32_t m = 1; m <= 1; m++) { |
| 4433 | GemmMicrokernelTester() |
| 4434 | .mr(1) |
| 4435 | .nr(8) |
| 4436 | .kr(2) |
| 4437 | .sr(1) |
| 4438 | .m(m) |
| 4439 | .n(n) |
| 4440 | .k(k) |
| 4441 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4442 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4443 | } |
| 4444 | } |
| 4445 | } |
| 4446 | } |
| 4447 | |
| 4448 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_div_8) { |
| 4449 | TEST_REQUIRES_ARM_NEON; |
| 4450 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4451 | for (size_t k = 1; k <= 80; k += 17) { |
| 4452 | GemmMicrokernelTester() |
| 4453 | .mr(1) |
| 4454 | .nr(8) |
| 4455 | .kr(2) |
| 4456 | .sr(1) |
| 4457 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4458 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4459 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4460 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4461 | } |
| 4462 | } |
| 4463 | } |
| 4464 | |
| 4465 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_div_8_strided_cn) { |
| 4466 | TEST_REQUIRES_ARM_NEON; |
| 4467 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4468 | for (size_t k = 1; k <= 80; k += 17) { |
| 4469 | GemmMicrokernelTester() |
| 4470 | .mr(1) |
| 4471 | .nr(8) |
| 4472 | .kr(2) |
| 4473 | .sr(1) |
| 4474 | .m(1) |
| 4475 | .n(n) |
| 4476 | .k(k) |
| 4477 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4478 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4479 | } |
| 4480 | } |
| 4481 | } |
| 4482 | |
| 4483 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_div_8_strided_a) { |
| 4484 | TEST_REQUIRES_ARM_NEON; |
| 4485 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4486 | for (size_t k = 1; k <= 80; k += 17) { |
| 4487 | GemmMicrokernelTester() |
| 4488 | .mr(1) |
| 4489 | .nr(8) |
| 4490 | .kr(2) |
| 4491 | .sr(1) |
| 4492 | .m(1) |
| 4493 | .n(n) |
| 4494 | .k(k) |
| 4495 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4496 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4497 | } |
| 4498 | } |
| 4499 | } |
| 4500 | |
| 4501 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, n_div_8_subtile) { |
| 4502 | TEST_REQUIRES_ARM_NEON; |
| 4503 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4504 | for (size_t k = 1; k <= 80; k += 17) { |
| 4505 | for (uint32_t m = 1; m <= 1; m++) { |
| 4506 | GemmMicrokernelTester() |
| 4507 | .mr(1) |
| 4508 | .nr(8) |
| 4509 | .kr(2) |
| 4510 | .sr(1) |
| 4511 | .m(m) |
| 4512 | .n(n) |
| 4513 | .k(k) |
| 4514 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4515 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4516 | } |
| 4517 | } |
| 4518 | } |
| 4519 | } |
| 4520 | |
| 4521 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, strided_cm_subtile) { |
| 4522 | TEST_REQUIRES_ARM_NEON; |
| 4523 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4524 | for (uint32_t n = 1; n <= 8; n++) { |
| 4525 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4526 | GemmMicrokernelTester() |
| 4527 | .mr(1) |
| 4528 | .nr(8) |
| 4529 | .kr(2) |
| 4530 | .sr(1) |
| 4531 | .m(m) |
| 4532 | .n(n) |
| 4533 | .k(k) |
| 4534 | .cm_stride(11) |
| 4535 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4536 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4537 | } |
| 4538 | } |
| 4539 | } |
| 4540 | } |
| 4541 | |
| 4542 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, qmin) { |
| 4543 | TEST_REQUIRES_ARM_NEON; |
| 4544 | GemmMicrokernelTester() |
| 4545 | .mr(1) |
| 4546 | .nr(8) |
| 4547 | .kr(2) |
| 4548 | .sr(1) |
| 4549 | .m(1) |
| 4550 | .n(8) |
| 4551 | .k(16) |
| 4552 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4553 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4554 | } |
| 4555 | |
| 4556 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, qmax) { |
| 4557 | TEST_REQUIRES_ARM_NEON; |
| 4558 | GemmMicrokernelTester() |
| 4559 | .mr(1) |
| 4560 | .nr(8) |
| 4561 | .kr(2) |
| 4562 | .sr(1) |
| 4563 | .m(1) |
| 4564 | .n(8) |
| 4565 | .k(16) |
| 4566 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4567 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4568 | } |
| 4569 | |
| 4570 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD2R, strided_cm) { |
| 4571 | TEST_REQUIRES_ARM_NEON; |
| 4572 | GemmMicrokernelTester() |
| 4573 | .mr(1) |
| 4574 | .nr(8) |
| 4575 | .kr(2) |
| 4576 | .sr(1) |
| 4577 | .m(1) |
| 4578 | .n(8) |
| 4579 | .k(16) |
| 4580 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4581 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4582 | } |
| 4583 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4584 | |
| 4585 | |
| 4586 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 4587 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8) { |
| 4588 | TEST_REQUIRES_ARM_NEON; |
| 4589 | GemmMicrokernelTester() |
| 4590 | .mr(1) |
| 4591 | .nr(8) |
| 4592 | .kr(4) |
| 4593 | .sr(2) |
| 4594 | .m(1) |
| 4595 | .n(8) |
| 4596 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4597 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4598 | } |
| 4599 | |
| 4600 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cn) { |
| 4601 | TEST_REQUIRES_ARM_NEON; |
| 4602 | GemmMicrokernelTester() |
| 4603 | .mr(1) |
| 4604 | .nr(8) |
| 4605 | .kr(4) |
| 4606 | .sr(2) |
| 4607 | .m(1) |
| 4608 | .n(8) |
| 4609 | .k(8) |
| 4610 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4611 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4612 | } |
| 4613 | |
| 4614 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_strided_a) { |
| 4615 | TEST_REQUIRES_ARM_NEON; |
| 4616 | GemmMicrokernelTester() |
| 4617 | .mr(1) |
| 4618 | .nr(8) |
| 4619 | .kr(4) |
| 4620 | .sr(2) |
| 4621 | .m(1) |
| 4622 | .n(8) |
| 4623 | .k(8) |
| 4624 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4625 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4626 | } |
| 4627 | |
| 4628 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile) { |
| 4629 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4630 | for (uint32_t n = 1; n <= 8; n++) { |
| 4631 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4632 | GemmMicrokernelTester() |
| 4633 | .mr(1) |
| 4634 | .nr(8) |
| 4635 | .kr(4) |
| 4636 | .sr(2) |
| 4637 | .m(m) |
| 4638 | .n(n) |
| 4639 | .k(8) |
| 4640 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4641 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4642 | } |
| 4643 | } |
| 4644 | } |
| 4645 | |
| 4646 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile_m) { |
| 4647 | TEST_REQUIRES_ARM_NEON; |
| 4648 | for (uint32_t m = 1; m <= 1; m++) { |
| 4649 | GemmMicrokernelTester() |
| 4650 | .mr(1) |
| 4651 | .nr(8) |
| 4652 | .kr(4) |
| 4653 | .sr(2) |
| 4654 | .m(m) |
| 4655 | .n(8) |
| 4656 | .k(8) |
| 4657 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4658 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4659 | } |
| 4660 | } |
| 4661 | |
| 4662 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile_n) { |
| 4663 | TEST_REQUIRES_ARM_NEON; |
| 4664 | for (uint32_t n = 1; n <= 8; n++) { |
| 4665 | GemmMicrokernelTester() |
| 4666 | .mr(1) |
| 4667 | .nr(8) |
| 4668 | .kr(4) |
| 4669 | .sr(2) |
| 4670 | .m(1) |
| 4671 | .n(n) |
| 4672 | .k(8) |
| 4673 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4674 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4675 | } |
| 4676 | } |
| 4677 | |
| 4678 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_lt_8) { |
| 4679 | TEST_REQUIRES_ARM_NEON; |
| 4680 | for (size_t k = 1; k < 8; k++) { |
| 4681 | GemmMicrokernelTester() |
| 4682 | .mr(1) |
| 4683 | .nr(8) |
| 4684 | .kr(4) |
| 4685 | .sr(2) |
| 4686 | .m(1) |
| 4687 | .n(8) |
| 4688 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4689 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4690 | } |
| 4691 | } |
| 4692 | |
| 4693 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_lt_8_strided_a) { |
| 4694 | TEST_REQUIRES_ARM_NEON; |
| 4695 | for (size_t k = 1; k < 8; k++) { |
| 4696 | GemmMicrokernelTester() |
| 4697 | .mr(1) |
| 4698 | .nr(8) |
| 4699 | .kr(4) |
| 4700 | .sr(2) |
| 4701 | .m(1) |
| 4702 | .n(8) |
| 4703 | .k(k) |
| 4704 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4705 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4706 | } |
| 4707 | } |
| 4708 | |
| 4709 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_lt_8_subtile) { |
| 4710 | TEST_REQUIRES_ARM_NEON; |
| 4711 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4712 | for (uint32_t n = 1; n <= 8; n++) { |
| 4713 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4714 | GemmMicrokernelTester() |
| 4715 | .mr(1) |
| 4716 | .nr(8) |
| 4717 | .kr(4) |
| 4718 | .sr(2) |
| 4719 | .m(m) |
| 4720 | .n(n) |
| 4721 | .k(k) |
| 4722 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4723 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4724 | } |
| 4725 | } |
| 4726 | } |
| 4727 | } |
| 4728 | |
| 4729 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_gt_8) { |
| 4730 | TEST_REQUIRES_ARM_NEON; |
| 4731 | for (size_t k = 9; k < 16; k++) { |
| 4732 | GemmMicrokernelTester() |
| 4733 | .mr(1) |
| 4734 | .nr(8) |
| 4735 | .kr(4) |
| 4736 | .sr(2) |
| 4737 | .m(1) |
| 4738 | .n(8) |
| 4739 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4740 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4741 | } |
| 4742 | } |
| 4743 | |
| 4744 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_gt_8_strided_a) { |
| 4745 | TEST_REQUIRES_ARM_NEON; |
| 4746 | for (size_t k = 9; k < 16; k++) { |
| 4747 | GemmMicrokernelTester() |
| 4748 | .mr(1) |
| 4749 | .nr(8) |
| 4750 | .kr(4) |
| 4751 | .sr(2) |
| 4752 | .m(1) |
| 4753 | .n(8) |
| 4754 | .k(k) |
| 4755 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4756 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4757 | } |
| 4758 | } |
| 4759 | |
| 4760 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_gt_8_subtile) { |
| 4761 | TEST_REQUIRES_ARM_NEON; |
| 4762 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4763 | for (uint32_t n = 1; n <= 8; n++) { |
| 4764 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4765 | GemmMicrokernelTester() |
| 4766 | .mr(1) |
| 4767 | .nr(8) |
| 4768 | .kr(4) |
| 4769 | .sr(2) |
| 4770 | .m(m) |
| 4771 | .n(n) |
| 4772 | .k(k) |
| 4773 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4774 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4775 | } |
| 4776 | } |
| 4777 | } |
| 4778 | } |
| 4779 | |
| 4780 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_div_8) { |
| 4781 | TEST_REQUIRES_ARM_NEON; |
| 4782 | for (size_t k = 16; k <= 80; k += 8) { |
| 4783 | GemmMicrokernelTester() |
| 4784 | .mr(1) |
| 4785 | .nr(8) |
| 4786 | .kr(4) |
| 4787 | .sr(2) |
| 4788 | .m(1) |
| 4789 | .n(8) |
| 4790 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4791 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4792 | } |
| 4793 | } |
| 4794 | |
| 4795 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_div_8_strided_a) { |
| 4796 | TEST_REQUIRES_ARM_NEON; |
| 4797 | for (size_t k = 16; k <= 80; k += 8) { |
| 4798 | GemmMicrokernelTester() |
| 4799 | .mr(1) |
| 4800 | .nr(8) |
| 4801 | .kr(4) |
| 4802 | .sr(2) |
| 4803 | .m(1) |
| 4804 | .n(8) |
| 4805 | .k(k) |
| 4806 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4807 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4808 | } |
| 4809 | } |
| 4810 | |
| 4811 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_div_8_subtile) { |
| 4812 | TEST_REQUIRES_ARM_NEON; |
| 4813 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4814 | for (uint32_t n = 1; n <= 8; n++) { |
| 4815 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4816 | GemmMicrokernelTester() |
| 4817 | .mr(1) |
| 4818 | .nr(8) |
| 4819 | .kr(4) |
| 4820 | .sr(2) |
| 4821 | .m(m) |
| 4822 | .n(n) |
| 4823 | .k(k) |
| 4824 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4825 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4826 | } |
| 4827 | } |
| 4828 | } |
| 4829 | } |
| 4830 | |
| 4831 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8) { |
| 4832 | TEST_REQUIRES_ARM_NEON; |
| 4833 | for (uint32_t n = 9; n < 16; n++) { |
| 4834 | for (size_t k = 1; k <= 40; k += 9) { |
| 4835 | GemmMicrokernelTester() |
| 4836 | .mr(1) |
| 4837 | .nr(8) |
| 4838 | .kr(4) |
| 4839 | .sr(2) |
| 4840 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4841 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4842 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4843 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4844 | } |
| 4845 | } |
| 4846 | } |
| 4847 | |
| 4848 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8_strided_cn) { |
| 4849 | TEST_REQUIRES_ARM_NEON; |
| 4850 | for (uint32_t n = 9; n < 16; n++) { |
| 4851 | for (size_t k = 1; k <= 40; k += 9) { |
| 4852 | GemmMicrokernelTester() |
| 4853 | .mr(1) |
| 4854 | .nr(8) |
| 4855 | .kr(4) |
| 4856 | .sr(2) |
| 4857 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4858 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4859 | .k(k) |
| 4860 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4861 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4862 | } |
| 4863 | } |
| 4864 | } |
| 4865 | |
| 4866 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8_strided_a) { |
| 4867 | TEST_REQUIRES_ARM_NEON; |
| 4868 | for (uint32_t n = 9; n < 16; n++) { |
| 4869 | for (size_t k = 1; k <= 40; k += 9) { |
| 4870 | GemmMicrokernelTester() |
| 4871 | .mr(1) |
| 4872 | .nr(8) |
| 4873 | .kr(4) |
| 4874 | .sr(2) |
| 4875 | .m(1) |
| 4876 | .n(n) |
| 4877 | .k(k) |
| 4878 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4879 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4880 | } |
| 4881 | } |
| 4882 | } |
| 4883 | |
| 4884 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8_subtile) { |
| 4885 | TEST_REQUIRES_ARM_NEON; |
| 4886 | for (uint32_t n = 9; n < 16; n++) { |
| 4887 | for (size_t k = 1; k <= 40; k += 9) { |
| 4888 | for (uint32_t m = 1; m <= 1; m++) { |
| 4889 | GemmMicrokernelTester() |
| 4890 | .mr(1) |
| 4891 | .nr(8) |
| 4892 | .kr(4) |
| 4893 | .sr(2) |
| 4894 | .m(m) |
| 4895 | .n(n) |
| 4896 | .k(k) |
| 4897 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4898 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4899 | } |
| 4900 | } |
| 4901 | } |
| 4902 | } |
| 4903 | |
| 4904 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8) { |
| 4905 | TEST_REQUIRES_ARM_NEON; |
| 4906 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4907 | for (size_t k = 1; k <= 40; k += 9) { |
| 4908 | GemmMicrokernelTester() |
| 4909 | .mr(1) |
| 4910 | .nr(8) |
| 4911 | .kr(4) |
| 4912 | .sr(2) |
| 4913 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 4914 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4915 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4916 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4917 | } |
| 4918 | } |
| 4919 | } |
| 4920 | |
| 4921 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8_strided_cn) { |
| 4922 | TEST_REQUIRES_ARM_NEON; |
| 4923 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4924 | for (size_t k = 1; k <= 40; k += 9) { |
| 4925 | GemmMicrokernelTester() |
| 4926 | .mr(1) |
| 4927 | .nr(8) |
| 4928 | .kr(4) |
| 4929 | .sr(2) |
| 4930 | .m(1) |
| 4931 | .n(n) |
| 4932 | .k(k) |
| 4933 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4934 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4935 | } |
| 4936 | } |
| 4937 | } |
| 4938 | |
| 4939 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8_strided_a) { |
| 4940 | TEST_REQUIRES_ARM_NEON; |
| 4941 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4942 | for (size_t k = 1; k <= 40; k += 9) { |
| 4943 | GemmMicrokernelTester() |
| 4944 | .mr(1) |
| 4945 | .nr(8) |
| 4946 | .kr(4) |
| 4947 | .sr(2) |
| 4948 | .m(1) |
| 4949 | .n(n) |
| 4950 | .k(k) |
| 4951 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4952 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4953 | } |
| 4954 | } |
| 4955 | } |
| 4956 | |
| 4957 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8_subtile) { |
| 4958 | TEST_REQUIRES_ARM_NEON; |
| 4959 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 4960 | for (size_t k = 1; k <= 40; k += 9) { |
| 4961 | for (uint32_t m = 1; m <= 1; m++) { |
| 4962 | GemmMicrokernelTester() |
| 4963 | .mr(1) |
| 4964 | .nr(8) |
| 4965 | .kr(4) |
| 4966 | .sr(2) |
| 4967 | .m(m) |
| 4968 | .n(n) |
| 4969 | .k(k) |
| 4970 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4971 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4972 | } |
| 4973 | } |
| 4974 | } |
| 4975 | } |
| 4976 | |
| 4977 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cm_subtile) { |
| 4978 | TEST_REQUIRES_ARM_NEON; |
| 4979 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 4980 | for (uint32_t n = 1; n <= 8; n++) { |
| 4981 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4982 | GemmMicrokernelTester() |
| 4983 | .mr(1) |
| 4984 | .nr(8) |
| 4985 | .kr(4) |
| 4986 | .sr(2) |
| 4987 | .m(m) |
| 4988 | .n(n) |
| 4989 | .k(k) |
| 4990 | .cm_stride(11) |
| 4991 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 4992 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 4993 | } |
| 4994 | } |
| 4995 | } |
| 4996 | } |
| 4997 | |
| 4998 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, qmin) { |
| 4999 | TEST_REQUIRES_ARM_NEON; |
| 5000 | GemmMicrokernelTester() |
| 5001 | .mr(1) |
| 5002 | .nr(8) |
| 5003 | .kr(4) |
| 5004 | .sr(2) |
| 5005 | .m(1) |
| 5006 | .n(8) |
| 5007 | .k(8) |
| 5008 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5009 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5010 | } |
| 5011 | |
| 5012 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, qmax) { |
| 5013 | TEST_REQUIRES_ARM_NEON; |
| 5014 | GemmMicrokernelTester() |
| 5015 | .mr(1) |
| 5016 | .nr(8) |
| 5017 | .kr(4) |
| 5018 | .sr(2) |
| 5019 | .m(1) |
| 5020 | .n(8) |
| 5021 | .k(8) |
| 5022 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5023 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5024 | } |
| 5025 | |
| 5026 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cm) { |
| 5027 | TEST_REQUIRES_ARM_NEON; |
| 5028 | GemmMicrokernelTester() |
| 5029 | .mr(1) |
| 5030 | .nr(8) |
| 5031 | .kr(4) |
| 5032 | .sr(2) |
| 5033 | .m(1) |
| 5034 | .n(8) |
| 5035 | .k(8) |
| 5036 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5037 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5038 | } |
| 5039 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 5040 | |
| 5041 | |
| 5042 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 5043 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_eq_16) { |
| 5044 | TEST_REQUIRES_ARM_NEON; |
| 5045 | GemmMicrokernelTester() |
| 5046 | .mr(3) |
| 5047 | .nr(8) |
| 5048 | .kr(4) |
| 5049 | .sr(2) |
| 5050 | .m(3) |
| 5051 | .n(8) |
| 5052 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5053 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5054 | } |
| 5055 | |
| 5056 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, strided_cn) { |
| 5057 | TEST_REQUIRES_ARM_NEON; |
| 5058 | GemmMicrokernelTester() |
| 5059 | .mr(3) |
| 5060 | .nr(8) |
| 5061 | .kr(4) |
| 5062 | .sr(2) |
| 5063 | .m(3) |
| 5064 | .n(8) |
| 5065 | .k(16) |
| 5066 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5067 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5068 | } |
| 5069 | |
| 5070 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_eq_16_strided_a) { |
| 5071 | TEST_REQUIRES_ARM_NEON; |
| 5072 | GemmMicrokernelTester() |
| 5073 | .mr(3) |
| 5074 | .nr(8) |
| 5075 | .kr(4) |
| 5076 | .sr(2) |
| 5077 | .m(3) |
| 5078 | .n(8) |
| 5079 | .k(16) |
| 5080 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5081 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5082 | } |
| 5083 | |
| 5084 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_eq_16_subtile) { |
| 5085 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5086 | for (uint32_t n = 1; n <= 8; n++) { |
| 5087 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5088 | GemmMicrokernelTester() |
| 5089 | .mr(3) |
| 5090 | .nr(8) |
| 5091 | .kr(4) |
| 5092 | .sr(2) |
| 5093 | .m(m) |
| 5094 | .n(n) |
| 5095 | .k(16) |
| 5096 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5097 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5098 | } |
| 5099 | } |
| 5100 | } |
| 5101 | |
| 5102 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_eq_16_subtile_m) { |
| 5103 | TEST_REQUIRES_ARM_NEON; |
| 5104 | for (uint32_t m = 1; m <= 3; m++) { |
| 5105 | GemmMicrokernelTester() |
| 5106 | .mr(3) |
| 5107 | .nr(8) |
| 5108 | .kr(4) |
| 5109 | .sr(2) |
| 5110 | .m(m) |
| 5111 | .n(8) |
| 5112 | .k(16) |
| 5113 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5114 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5115 | } |
| 5116 | } |
| 5117 | |
| 5118 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_eq_16_subtile_n) { |
| 5119 | TEST_REQUIRES_ARM_NEON; |
| 5120 | for (uint32_t n = 1; n <= 8; n++) { |
| 5121 | GemmMicrokernelTester() |
| 5122 | .mr(3) |
| 5123 | .nr(8) |
| 5124 | .kr(4) |
| 5125 | .sr(2) |
| 5126 | .m(3) |
| 5127 | .n(n) |
| 5128 | .k(16) |
| 5129 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5130 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5131 | } |
| 5132 | } |
| 5133 | |
| 5134 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_lt_16) { |
| 5135 | TEST_REQUIRES_ARM_NEON; |
| 5136 | for (size_t k = 1; k < 16; k++) { |
| 5137 | GemmMicrokernelTester() |
| 5138 | .mr(3) |
| 5139 | .nr(8) |
| 5140 | .kr(4) |
| 5141 | .sr(2) |
| 5142 | .m(3) |
| 5143 | .n(8) |
| 5144 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5145 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5146 | } |
| 5147 | } |
| 5148 | |
| 5149 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_lt_16_strided_a) { |
| 5150 | TEST_REQUIRES_ARM_NEON; |
| 5151 | for (size_t k = 1; k < 16; k++) { |
| 5152 | GemmMicrokernelTester() |
| 5153 | .mr(3) |
| 5154 | .nr(8) |
| 5155 | .kr(4) |
| 5156 | .sr(2) |
| 5157 | .m(3) |
| 5158 | .n(8) |
| 5159 | .k(k) |
| 5160 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5161 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5162 | } |
| 5163 | } |
| 5164 | |
| 5165 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_lt_16_subtile) { |
| 5166 | TEST_REQUIRES_ARM_NEON; |
| 5167 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5168 | for (uint32_t n = 1; n <= 8; n++) { |
| 5169 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5170 | GemmMicrokernelTester() |
| 5171 | .mr(3) |
| 5172 | .nr(8) |
| 5173 | .kr(4) |
| 5174 | .sr(2) |
| 5175 | .m(m) |
| 5176 | .n(n) |
| 5177 | .k(k) |
| 5178 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5179 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5180 | } |
| 5181 | } |
| 5182 | } |
| 5183 | } |
| 5184 | |
| 5185 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_gt_16) { |
| 5186 | TEST_REQUIRES_ARM_NEON; |
| 5187 | for (size_t k = 17; k < 32; k++) { |
| 5188 | GemmMicrokernelTester() |
| 5189 | .mr(3) |
| 5190 | .nr(8) |
| 5191 | .kr(4) |
| 5192 | .sr(2) |
| 5193 | .m(3) |
| 5194 | .n(8) |
| 5195 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5196 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5197 | } |
| 5198 | } |
| 5199 | |
| 5200 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_gt_16_strided_a) { |
| 5201 | TEST_REQUIRES_ARM_NEON; |
| 5202 | for (size_t k = 17; k < 32; k++) { |
| 5203 | GemmMicrokernelTester() |
| 5204 | .mr(3) |
| 5205 | .nr(8) |
| 5206 | .kr(4) |
| 5207 | .sr(2) |
| 5208 | .m(3) |
| 5209 | .n(8) |
| 5210 | .k(k) |
| 5211 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5212 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5213 | } |
| 5214 | } |
| 5215 | |
| 5216 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_gt_16_subtile) { |
| 5217 | TEST_REQUIRES_ARM_NEON; |
| 5218 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5219 | for (uint32_t n = 1; n <= 8; n++) { |
| 5220 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5221 | GemmMicrokernelTester() |
| 5222 | .mr(3) |
| 5223 | .nr(8) |
| 5224 | .kr(4) |
| 5225 | .sr(2) |
| 5226 | .m(m) |
| 5227 | .n(n) |
| 5228 | .k(k) |
| 5229 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5230 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5231 | } |
| 5232 | } |
| 5233 | } |
| 5234 | } |
| 5235 | |
| 5236 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_div_16) { |
| 5237 | TEST_REQUIRES_ARM_NEON; |
| 5238 | for (size_t k = 32; k <= 160; k += 16) { |
| 5239 | GemmMicrokernelTester() |
| 5240 | .mr(3) |
| 5241 | .nr(8) |
| 5242 | .kr(4) |
| 5243 | .sr(2) |
| 5244 | .m(3) |
| 5245 | .n(8) |
| 5246 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5247 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5248 | } |
| 5249 | } |
| 5250 | |
| 5251 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_div_16_strided_a) { |
| 5252 | TEST_REQUIRES_ARM_NEON; |
| 5253 | for (size_t k = 32; k <= 160; k += 16) { |
| 5254 | GemmMicrokernelTester() |
| 5255 | .mr(3) |
| 5256 | .nr(8) |
| 5257 | .kr(4) |
| 5258 | .sr(2) |
| 5259 | .m(3) |
| 5260 | .n(8) |
| 5261 | .k(k) |
| 5262 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5263 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5264 | } |
| 5265 | } |
| 5266 | |
| 5267 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, k_div_16_subtile) { |
| 5268 | TEST_REQUIRES_ARM_NEON; |
| 5269 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5270 | for (uint32_t n = 1; n <= 8; n++) { |
| 5271 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5272 | GemmMicrokernelTester() |
| 5273 | .mr(3) |
| 5274 | .nr(8) |
| 5275 | .kr(4) |
| 5276 | .sr(2) |
| 5277 | .m(m) |
| 5278 | .n(n) |
| 5279 | .k(k) |
| 5280 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5281 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5282 | } |
| 5283 | } |
| 5284 | } |
| 5285 | } |
| 5286 | |
| 5287 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_gt_8) { |
| 5288 | TEST_REQUIRES_ARM_NEON; |
| 5289 | for (uint32_t n = 9; n < 16; n++) { |
| 5290 | for (size_t k = 1; k <= 80; k += 17) { |
| 5291 | GemmMicrokernelTester() |
| 5292 | .mr(3) |
| 5293 | .nr(8) |
| 5294 | .kr(4) |
| 5295 | .sr(2) |
| 5296 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5297 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5298 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5299 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5300 | } |
| 5301 | } |
| 5302 | } |
| 5303 | |
| 5304 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_gt_8_strided_cn) { |
| 5305 | TEST_REQUIRES_ARM_NEON; |
| 5306 | for (uint32_t n = 9; n < 16; n++) { |
| 5307 | for (size_t k = 1; k <= 80; k += 17) { |
| 5308 | GemmMicrokernelTester() |
| 5309 | .mr(3) |
| 5310 | .nr(8) |
| 5311 | .kr(4) |
| 5312 | .sr(2) |
| 5313 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5314 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5315 | .k(k) |
| 5316 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5317 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5318 | } |
| 5319 | } |
| 5320 | } |
| 5321 | |
| 5322 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_gt_8_strided_a) { |
| 5323 | TEST_REQUIRES_ARM_NEON; |
| 5324 | for (uint32_t n = 9; n < 16; n++) { |
| 5325 | for (size_t k = 1; k <= 80; k += 17) { |
| 5326 | GemmMicrokernelTester() |
| 5327 | .mr(3) |
| 5328 | .nr(8) |
| 5329 | .kr(4) |
| 5330 | .sr(2) |
| 5331 | .m(3) |
| 5332 | .n(n) |
| 5333 | .k(k) |
| 5334 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5335 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5336 | } |
| 5337 | } |
| 5338 | } |
| 5339 | |
| 5340 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_gt_8_subtile) { |
| 5341 | TEST_REQUIRES_ARM_NEON; |
| 5342 | for (uint32_t n = 9; n < 16; n++) { |
| 5343 | for (size_t k = 1; k <= 80; k += 17) { |
| 5344 | for (uint32_t m = 1; m <= 3; m++) { |
| 5345 | GemmMicrokernelTester() |
| 5346 | .mr(3) |
| 5347 | .nr(8) |
| 5348 | .kr(4) |
| 5349 | .sr(2) |
| 5350 | .m(m) |
| 5351 | .n(n) |
| 5352 | .k(k) |
| 5353 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5354 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5355 | } |
| 5356 | } |
| 5357 | } |
| 5358 | } |
| 5359 | |
| 5360 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_div_8) { |
| 5361 | TEST_REQUIRES_ARM_NEON; |
| 5362 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 5363 | for (size_t k = 1; k <= 80; k += 17) { |
| 5364 | GemmMicrokernelTester() |
| 5365 | .mr(3) |
| 5366 | .nr(8) |
| 5367 | .kr(4) |
| 5368 | .sr(2) |
| 5369 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5370 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5371 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5372 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5373 | } |
| 5374 | } |
| 5375 | } |
| 5376 | |
| 5377 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_div_8_strided_cn) { |
| 5378 | TEST_REQUIRES_ARM_NEON; |
| 5379 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 5380 | for (size_t k = 1; k <= 80; k += 17) { |
| 5381 | GemmMicrokernelTester() |
| 5382 | .mr(3) |
| 5383 | .nr(8) |
| 5384 | .kr(4) |
| 5385 | .sr(2) |
| 5386 | .m(3) |
| 5387 | .n(n) |
| 5388 | .k(k) |
| 5389 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5390 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5391 | } |
| 5392 | } |
| 5393 | } |
| 5394 | |
| 5395 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_div_8_strided_a) { |
| 5396 | TEST_REQUIRES_ARM_NEON; |
| 5397 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 5398 | for (size_t k = 1; k <= 80; k += 17) { |
| 5399 | GemmMicrokernelTester() |
| 5400 | .mr(3) |
| 5401 | .nr(8) |
| 5402 | .kr(4) |
| 5403 | .sr(2) |
| 5404 | .m(3) |
| 5405 | .n(n) |
| 5406 | .k(k) |
| 5407 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5408 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5409 | } |
| 5410 | } |
| 5411 | } |
| 5412 | |
| 5413 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, n_div_8_subtile) { |
| 5414 | TEST_REQUIRES_ARM_NEON; |
| 5415 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 5416 | for (size_t k = 1; k <= 80; k += 17) { |
| 5417 | for (uint32_t m = 1; m <= 3; m++) { |
| 5418 | GemmMicrokernelTester() |
| 5419 | .mr(3) |
| 5420 | .nr(8) |
| 5421 | .kr(4) |
| 5422 | .sr(2) |
| 5423 | .m(m) |
| 5424 | .n(n) |
| 5425 | .k(k) |
| 5426 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5427 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5428 | } |
| 5429 | } |
| 5430 | } |
| 5431 | } |
| 5432 | |
| 5433 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, strided_cm_subtile) { |
| 5434 | TEST_REQUIRES_ARM_NEON; |
| 5435 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5436 | for (uint32_t n = 1; n <= 8; n++) { |
| 5437 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5438 | GemmMicrokernelTester() |
| 5439 | .mr(3) |
| 5440 | .nr(8) |
| 5441 | .kr(4) |
| 5442 | .sr(2) |
| 5443 | .m(m) |
| 5444 | .n(n) |
| 5445 | .k(k) |
| 5446 | .cm_stride(11) |
| 5447 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5448 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5449 | } |
| 5450 | } |
| 5451 | } |
| 5452 | } |
| 5453 | |
| 5454 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, qmin) { |
| 5455 | TEST_REQUIRES_ARM_NEON; |
| 5456 | GemmMicrokernelTester() |
| 5457 | .mr(3) |
| 5458 | .nr(8) |
| 5459 | .kr(4) |
| 5460 | .sr(2) |
| 5461 | .m(3) |
| 5462 | .n(8) |
| 5463 | .k(16) |
| 5464 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5465 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5466 | } |
| 5467 | |
| 5468 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, qmax) { |
| 5469 | TEST_REQUIRES_ARM_NEON; |
| 5470 | GemmMicrokernelTester() |
| 5471 | .mr(3) |
| 5472 | .nr(8) |
| 5473 | .kr(4) |
| 5474 | .sr(2) |
| 5475 | .m(3) |
| 5476 | .n(8) |
| 5477 | .k(16) |
| 5478 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5479 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5480 | } |
| 5481 | |
| 5482 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4S2__NEON_MLAL, strided_cm) { |
| 5483 | TEST_REQUIRES_ARM_NEON; |
| 5484 | GemmMicrokernelTester() |
| 5485 | .mr(3) |
| 5486 | .nr(8) |
| 5487 | .kr(4) |
| 5488 | .sr(2) |
| 5489 | .m(3) |
| 5490 | .n(8) |
| 5491 | .k(16) |
| 5492 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5493 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5494 | } |
| 5495 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 5496 | |
| 5497 | |
| 5498 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 5499 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_eq_16) { |
| 5500 | TEST_REQUIRES_ARM_NEON; |
| 5501 | GemmMicrokernelTester() |
| 5502 | .mr(3) |
| 5503 | .nr(16) |
| 5504 | .kr(4) |
| 5505 | .sr(2) |
| 5506 | .m(3) |
| 5507 | .n(16) |
| 5508 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5509 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5510 | } |
| 5511 | |
| 5512 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, strided_cn) { |
| 5513 | TEST_REQUIRES_ARM_NEON; |
| 5514 | GemmMicrokernelTester() |
| 5515 | .mr(3) |
| 5516 | .nr(16) |
| 5517 | .kr(4) |
| 5518 | .sr(2) |
| 5519 | .m(3) |
| 5520 | .n(16) |
| 5521 | .k(16) |
| 5522 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5523 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5524 | } |
| 5525 | |
| 5526 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_eq_16_strided_a) { |
| 5527 | TEST_REQUIRES_ARM_NEON; |
| 5528 | GemmMicrokernelTester() |
| 5529 | .mr(3) |
| 5530 | .nr(16) |
| 5531 | .kr(4) |
| 5532 | .sr(2) |
| 5533 | .m(3) |
| 5534 | .n(16) |
| 5535 | .k(16) |
| 5536 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5537 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5538 | } |
| 5539 | |
| 5540 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_eq_16_subtile) { |
| 5541 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5542 | for (uint32_t n = 1; n <= 16; n++) { |
| 5543 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5544 | GemmMicrokernelTester() |
| 5545 | .mr(3) |
| 5546 | .nr(16) |
| 5547 | .kr(4) |
| 5548 | .sr(2) |
| 5549 | .m(m) |
| 5550 | .n(n) |
| 5551 | .k(16) |
| 5552 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5553 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5554 | } |
| 5555 | } |
| 5556 | } |
| 5557 | |
| 5558 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_eq_16_subtile_m) { |
| 5559 | TEST_REQUIRES_ARM_NEON; |
| 5560 | for (uint32_t m = 1; m <= 3; m++) { |
| 5561 | GemmMicrokernelTester() |
| 5562 | .mr(3) |
| 5563 | .nr(16) |
| 5564 | .kr(4) |
| 5565 | .sr(2) |
| 5566 | .m(m) |
| 5567 | .n(16) |
| 5568 | .k(16) |
| 5569 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5570 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5571 | } |
| 5572 | } |
| 5573 | |
| 5574 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_eq_16_subtile_n) { |
| 5575 | TEST_REQUIRES_ARM_NEON; |
| 5576 | for (uint32_t n = 1; n <= 16; n++) { |
| 5577 | GemmMicrokernelTester() |
| 5578 | .mr(3) |
| 5579 | .nr(16) |
| 5580 | .kr(4) |
| 5581 | .sr(2) |
| 5582 | .m(3) |
| 5583 | .n(n) |
| 5584 | .k(16) |
| 5585 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5586 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5587 | } |
| 5588 | } |
| 5589 | |
| 5590 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_lt_16) { |
| 5591 | TEST_REQUIRES_ARM_NEON; |
| 5592 | for (size_t k = 1; k < 16; k++) { |
| 5593 | GemmMicrokernelTester() |
| 5594 | .mr(3) |
| 5595 | .nr(16) |
| 5596 | .kr(4) |
| 5597 | .sr(2) |
| 5598 | .m(3) |
| 5599 | .n(16) |
| 5600 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5601 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5602 | } |
| 5603 | } |
| 5604 | |
| 5605 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_lt_16_strided_a) { |
| 5606 | TEST_REQUIRES_ARM_NEON; |
| 5607 | for (size_t k = 1; k < 16; k++) { |
| 5608 | GemmMicrokernelTester() |
| 5609 | .mr(3) |
| 5610 | .nr(16) |
| 5611 | .kr(4) |
| 5612 | .sr(2) |
| 5613 | .m(3) |
| 5614 | .n(16) |
| 5615 | .k(k) |
| 5616 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5617 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5618 | } |
| 5619 | } |
| 5620 | |
| 5621 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_lt_16_subtile) { |
| 5622 | TEST_REQUIRES_ARM_NEON; |
| 5623 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5624 | for (uint32_t n = 1; n <= 16; n++) { |
| 5625 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5626 | GemmMicrokernelTester() |
| 5627 | .mr(3) |
| 5628 | .nr(16) |
| 5629 | .kr(4) |
| 5630 | .sr(2) |
| 5631 | .m(m) |
| 5632 | .n(n) |
| 5633 | .k(k) |
| 5634 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5635 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5636 | } |
| 5637 | } |
| 5638 | } |
| 5639 | } |
| 5640 | |
| 5641 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_gt_16) { |
| 5642 | TEST_REQUIRES_ARM_NEON; |
| 5643 | for (size_t k = 17; k < 32; k++) { |
| 5644 | GemmMicrokernelTester() |
| 5645 | .mr(3) |
| 5646 | .nr(16) |
| 5647 | .kr(4) |
| 5648 | .sr(2) |
| 5649 | .m(3) |
| 5650 | .n(16) |
| 5651 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5652 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5653 | } |
| 5654 | } |
| 5655 | |
| 5656 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_gt_16_strided_a) { |
| 5657 | TEST_REQUIRES_ARM_NEON; |
| 5658 | for (size_t k = 17; k < 32; k++) { |
| 5659 | GemmMicrokernelTester() |
| 5660 | .mr(3) |
| 5661 | .nr(16) |
| 5662 | .kr(4) |
| 5663 | .sr(2) |
| 5664 | .m(3) |
| 5665 | .n(16) |
| 5666 | .k(k) |
| 5667 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5668 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5669 | } |
| 5670 | } |
| 5671 | |
| 5672 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_gt_16_subtile) { |
| 5673 | TEST_REQUIRES_ARM_NEON; |
| 5674 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5675 | for (uint32_t n = 1; n <= 16; n++) { |
| 5676 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5677 | GemmMicrokernelTester() |
| 5678 | .mr(3) |
| 5679 | .nr(16) |
| 5680 | .kr(4) |
| 5681 | .sr(2) |
| 5682 | .m(m) |
| 5683 | .n(n) |
| 5684 | .k(k) |
| 5685 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5686 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5687 | } |
| 5688 | } |
| 5689 | } |
| 5690 | } |
| 5691 | |
| 5692 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_div_16) { |
| 5693 | TEST_REQUIRES_ARM_NEON; |
| 5694 | for (size_t k = 32; k <= 160; k += 16) { |
| 5695 | GemmMicrokernelTester() |
| 5696 | .mr(3) |
| 5697 | .nr(16) |
| 5698 | .kr(4) |
| 5699 | .sr(2) |
| 5700 | .m(3) |
| 5701 | .n(16) |
| 5702 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5703 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5704 | } |
| 5705 | } |
| 5706 | |
| 5707 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_div_16_strided_a) { |
| 5708 | TEST_REQUIRES_ARM_NEON; |
| 5709 | for (size_t k = 32; k <= 160; k += 16) { |
| 5710 | GemmMicrokernelTester() |
| 5711 | .mr(3) |
| 5712 | .nr(16) |
| 5713 | .kr(4) |
| 5714 | .sr(2) |
| 5715 | .m(3) |
| 5716 | .n(16) |
| 5717 | .k(k) |
| 5718 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5719 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5720 | } |
| 5721 | } |
| 5722 | |
| 5723 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, k_div_16_subtile) { |
| 5724 | TEST_REQUIRES_ARM_NEON; |
| 5725 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5726 | for (uint32_t n = 1; n <= 16; n++) { |
| 5727 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5728 | GemmMicrokernelTester() |
| 5729 | .mr(3) |
| 5730 | .nr(16) |
| 5731 | .kr(4) |
| 5732 | .sr(2) |
| 5733 | .m(m) |
| 5734 | .n(n) |
| 5735 | .k(k) |
| 5736 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5737 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5738 | } |
| 5739 | } |
| 5740 | } |
| 5741 | } |
| 5742 | |
| 5743 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_gt_16) { |
| 5744 | TEST_REQUIRES_ARM_NEON; |
| 5745 | for (uint32_t n = 17; n < 32; n++) { |
| 5746 | for (size_t k = 1; k <= 80; k += 17) { |
| 5747 | GemmMicrokernelTester() |
| 5748 | .mr(3) |
| 5749 | .nr(16) |
| 5750 | .kr(4) |
| 5751 | .sr(2) |
| 5752 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5753 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5754 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5755 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5756 | } |
| 5757 | } |
| 5758 | } |
| 5759 | |
| 5760 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_gt_16_strided_cn) { |
| 5761 | TEST_REQUIRES_ARM_NEON; |
| 5762 | for (uint32_t n = 17; n < 32; n++) { |
| 5763 | for (size_t k = 1; k <= 80; k += 17) { |
| 5764 | GemmMicrokernelTester() |
| 5765 | .mr(3) |
| 5766 | .nr(16) |
| 5767 | .kr(4) |
| 5768 | .sr(2) |
| 5769 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5770 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5771 | .k(k) |
| 5772 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5773 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5774 | } |
| 5775 | } |
| 5776 | } |
| 5777 | |
| 5778 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_gt_16_strided_a) { |
| 5779 | TEST_REQUIRES_ARM_NEON; |
| 5780 | for (uint32_t n = 17; n < 32; n++) { |
| 5781 | for (size_t k = 1; k <= 80; k += 17) { |
| 5782 | GemmMicrokernelTester() |
| 5783 | .mr(3) |
| 5784 | .nr(16) |
| 5785 | .kr(4) |
| 5786 | .sr(2) |
| 5787 | .m(3) |
| 5788 | .n(n) |
| 5789 | .k(k) |
| 5790 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5791 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5792 | } |
| 5793 | } |
| 5794 | } |
| 5795 | |
| 5796 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_gt_16_subtile) { |
| 5797 | TEST_REQUIRES_ARM_NEON; |
| 5798 | for (uint32_t n = 17; n < 32; n++) { |
| 5799 | for (size_t k = 1; k <= 80; k += 17) { |
| 5800 | for (uint32_t m = 1; m <= 3; m++) { |
| 5801 | GemmMicrokernelTester() |
| 5802 | .mr(3) |
| 5803 | .nr(16) |
| 5804 | .kr(4) |
| 5805 | .sr(2) |
| 5806 | .m(m) |
| 5807 | .n(n) |
| 5808 | .k(k) |
| 5809 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5810 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5811 | } |
| 5812 | } |
| 5813 | } |
| 5814 | } |
| 5815 | |
| 5816 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_div_16) { |
| 5817 | TEST_REQUIRES_ARM_NEON; |
| 5818 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5819 | for (size_t k = 1; k <= 80; k += 17) { |
| 5820 | GemmMicrokernelTester() |
| 5821 | .mr(3) |
| 5822 | .nr(16) |
| 5823 | .kr(4) |
| 5824 | .sr(2) |
| 5825 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 5826 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5827 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5828 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5829 | } |
| 5830 | } |
| 5831 | } |
| 5832 | |
| 5833 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_div_16_strided_cn) { |
| 5834 | TEST_REQUIRES_ARM_NEON; |
| 5835 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5836 | for (size_t k = 1; k <= 80; k += 17) { |
| 5837 | GemmMicrokernelTester() |
| 5838 | .mr(3) |
| 5839 | .nr(16) |
| 5840 | .kr(4) |
| 5841 | .sr(2) |
| 5842 | .m(3) |
| 5843 | .n(n) |
| 5844 | .k(k) |
| 5845 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5846 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5847 | } |
| 5848 | } |
| 5849 | } |
| 5850 | |
| 5851 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_div_16_strided_a) { |
| 5852 | TEST_REQUIRES_ARM_NEON; |
| 5853 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5854 | for (size_t k = 1; k <= 80; k += 17) { |
| 5855 | GemmMicrokernelTester() |
| 5856 | .mr(3) |
| 5857 | .nr(16) |
| 5858 | .kr(4) |
| 5859 | .sr(2) |
| 5860 | .m(3) |
| 5861 | .n(n) |
| 5862 | .k(k) |
| 5863 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5864 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5865 | } |
| 5866 | } |
| 5867 | } |
| 5868 | |
| 5869 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, n_div_16_subtile) { |
| 5870 | TEST_REQUIRES_ARM_NEON; |
| 5871 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 5872 | for (size_t k = 1; k <= 80; k += 17) { |
| 5873 | for (uint32_t m = 1; m <= 3; m++) { |
| 5874 | GemmMicrokernelTester() |
| 5875 | .mr(3) |
| 5876 | .nr(16) |
| 5877 | .kr(4) |
| 5878 | .sr(2) |
| 5879 | .m(m) |
| 5880 | .n(n) |
| 5881 | .k(k) |
| 5882 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5883 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5884 | } |
| 5885 | } |
| 5886 | } |
| 5887 | } |
| 5888 | |
| 5889 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, strided_cm_subtile) { |
| 5890 | TEST_REQUIRES_ARM_NEON; |
| 5891 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5892 | for (uint32_t n = 1; n <= 16; n++) { |
| 5893 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5894 | GemmMicrokernelTester() |
| 5895 | .mr(3) |
| 5896 | .nr(16) |
| 5897 | .kr(4) |
| 5898 | .sr(2) |
| 5899 | .m(m) |
| 5900 | .n(n) |
| 5901 | .k(k) |
| 5902 | .cm_stride(19) |
| 5903 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5904 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5905 | } |
| 5906 | } |
| 5907 | } |
| 5908 | } |
| 5909 | |
| 5910 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, qmin) { |
| 5911 | TEST_REQUIRES_ARM_NEON; |
| 5912 | GemmMicrokernelTester() |
| 5913 | .mr(3) |
| 5914 | .nr(16) |
| 5915 | .kr(4) |
| 5916 | .sr(2) |
| 5917 | .m(3) |
| 5918 | .n(16) |
| 5919 | .k(16) |
| 5920 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5921 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5922 | } |
| 5923 | |
| 5924 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, qmax) { |
| 5925 | TEST_REQUIRES_ARM_NEON; |
| 5926 | GemmMicrokernelTester() |
| 5927 | .mr(3) |
| 5928 | .nr(16) |
| 5929 | .kr(4) |
| 5930 | .sr(2) |
| 5931 | .m(3) |
| 5932 | .n(16) |
| 5933 | .k(16) |
| 5934 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5935 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5936 | } |
| 5937 | |
| 5938 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4S2__NEON_MLAL, strided_cm) { |
| 5939 | TEST_REQUIRES_ARM_NEON; |
| 5940 | GemmMicrokernelTester() |
| 5941 | .mr(3) |
| 5942 | .nr(16) |
| 5943 | .kr(4) |
| 5944 | .sr(2) |
| 5945 | .m(3) |
| 5946 | .n(16) |
| 5947 | .k(16) |
| 5948 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5949 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5950 | } |
| 5951 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 5952 | |
| 5953 | |
| 5954 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 5955 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16) { |
| 5956 | TEST_REQUIRES_ARM_NEON; |
| 5957 | GemmMicrokernelTester() |
| 5958 | .mr(4) |
| 5959 | .nr(16) |
| 5960 | .kr(4) |
| 5961 | .sr(2) |
| 5962 | .m(4) |
| 5963 | .n(16) |
| 5964 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5965 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5966 | } |
| 5967 | |
| 5968 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, strided_cn) { |
| 5969 | TEST_REQUIRES_ARM_NEON; |
| 5970 | GemmMicrokernelTester() |
| 5971 | .mr(4) |
| 5972 | .nr(16) |
| 5973 | .kr(4) |
| 5974 | .sr(2) |
| 5975 | .m(4) |
| 5976 | .n(16) |
| 5977 | .k(16) |
| 5978 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5979 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5980 | } |
| 5981 | |
| 5982 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16_strided_a) { |
| 5983 | TEST_REQUIRES_ARM_NEON; |
| 5984 | GemmMicrokernelTester() |
| 5985 | .mr(4) |
| 5986 | .nr(16) |
| 5987 | .kr(4) |
| 5988 | .sr(2) |
| 5989 | .m(4) |
| 5990 | .n(16) |
| 5991 | .k(16) |
| 5992 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 5993 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 5994 | } |
| 5995 | |
| 5996 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16_subtile) { |
| 5997 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 5998 | for (uint32_t n = 1; n <= 16; n++) { |
| 5999 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6000 | GemmMicrokernelTester() |
| 6001 | .mr(4) |
| 6002 | .nr(16) |
| 6003 | .kr(4) |
| 6004 | .sr(2) |
| 6005 | .m(m) |
| 6006 | .n(n) |
| 6007 | .k(16) |
| 6008 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6009 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6010 | } |
| 6011 | } |
| 6012 | } |
| 6013 | |
| 6014 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16_subtile_m) { |
| 6015 | TEST_REQUIRES_ARM_NEON; |
| 6016 | for (uint32_t m = 1; m <= 4; m++) { |
| 6017 | GemmMicrokernelTester() |
| 6018 | .mr(4) |
| 6019 | .nr(16) |
| 6020 | .kr(4) |
| 6021 | .sr(2) |
| 6022 | .m(m) |
| 6023 | .n(16) |
| 6024 | .k(16) |
| 6025 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6026 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6027 | } |
| 6028 | } |
| 6029 | |
| 6030 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_eq_16_subtile_n) { |
| 6031 | TEST_REQUIRES_ARM_NEON; |
| 6032 | for (uint32_t n = 1; n <= 16; n++) { |
| 6033 | GemmMicrokernelTester() |
| 6034 | .mr(4) |
| 6035 | .nr(16) |
| 6036 | .kr(4) |
| 6037 | .sr(2) |
| 6038 | .m(4) |
| 6039 | .n(n) |
| 6040 | .k(16) |
| 6041 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6042 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6043 | } |
| 6044 | } |
| 6045 | |
| 6046 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_lt_16) { |
| 6047 | TEST_REQUIRES_ARM_NEON; |
| 6048 | for (size_t k = 1; k < 16; k++) { |
| 6049 | GemmMicrokernelTester() |
| 6050 | .mr(4) |
| 6051 | .nr(16) |
| 6052 | .kr(4) |
| 6053 | .sr(2) |
| 6054 | .m(4) |
| 6055 | .n(16) |
| 6056 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6057 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6058 | } |
| 6059 | } |
| 6060 | |
| 6061 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_lt_16_strided_a) { |
| 6062 | TEST_REQUIRES_ARM_NEON; |
| 6063 | for (size_t k = 1; k < 16; k++) { |
| 6064 | GemmMicrokernelTester() |
| 6065 | .mr(4) |
| 6066 | .nr(16) |
| 6067 | .kr(4) |
| 6068 | .sr(2) |
| 6069 | .m(4) |
| 6070 | .n(16) |
| 6071 | .k(k) |
| 6072 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6073 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6074 | } |
| 6075 | } |
| 6076 | |
| 6077 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_lt_16_subtile) { |
| 6078 | TEST_REQUIRES_ARM_NEON; |
| 6079 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6080 | for (uint32_t n = 1; n <= 16; n++) { |
| 6081 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6082 | GemmMicrokernelTester() |
| 6083 | .mr(4) |
| 6084 | .nr(16) |
| 6085 | .kr(4) |
| 6086 | .sr(2) |
| 6087 | .m(m) |
| 6088 | .n(n) |
| 6089 | .k(k) |
| 6090 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6091 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6092 | } |
| 6093 | } |
| 6094 | } |
| 6095 | } |
| 6096 | |
| 6097 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_gt_16) { |
| 6098 | TEST_REQUIRES_ARM_NEON; |
| 6099 | for (size_t k = 17; k < 32; k++) { |
| 6100 | GemmMicrokernelTester() |
| 6101 | .mr(4) |
| 6102 | .nr(16) |
| 6103 | .kr(4) |
| 6104 | .sr(2) |
| 6105 | .m(4) |
| 6106 | .n(16) |
| 6107 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6108 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6109 | } |
| 6110 | } |
| 6111 | |
| 6112 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_gt_16_strided_a) { |
| 6113 | TEST_REQUIRES_ARM_NEON; |
| 6114 | for (size_t k = 17; k < 32; k++) { |
| 6115 | GemmMicrokernelTester() |
| 6116 | .mr(4) |
| 6117 | .nr(16) |
| 6118 | .kr(4) |
| 6119 | .sr(2) |
| 6120 | .m(4) |
| 6121 | .n(16) |
| 6122 | .k(k) |
| 6123 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6124 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6125 | } |
| 6126 | } |
| 6127 | |
| 6128 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_gt_16_subtile) { |
| 6129 | TEST_REQUIRES_ARM_NEON; |
| 6130 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6131 | for (uint32_t n = 1; n <= 16; n++) { |
| 6132 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6133 | GemmMicrokernelTester() |
| 6134 | .mr(4) |
| 6135 | .nr(16) |
| 6136 | .kr(4) |
| 6137 | .sr(2) |
| 6138 | .m(m) |
| 6139 | .n(n) |
| 6140 | .k(k) |
| 6141 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6142 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6143 | } |
| 6144 | } |
| 6145 | } |
| 6146 | } |
| 6147 | |
| 6148 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_div_16) { |
| 6149 | TEST_REQUIRES_ARM_NEON; |
| 6150 | for (size_t k = 32; k <= 160; k += 16) { |
| 6151 | GemmMicrokernelTester() |
| 6152 | .mr(4) |
| 6153 | .nr(16) |
| 6154 | .kr(4) |
| 6155 | .sr(2) |
| 6156 | .m(4) |
| 6157 | .n(16) |
| 6158 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6159 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6160 | } |
| 6161 | } |
| 6162 | |
| 6163 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_div_16_strided_a) { |
| 6164 | TEST_REQUIRES_ARM_NEON; |
| 6165 | for (size_t k = 32; k <= 160; k += 16) { |
| 6166 | GemmMicrokernelTester() |
| 6167 | .mr(4) |
| 6168 | .nr(16) |
| 6169 | .kr(4) |
| 6170 | .sr(2) |
| 6171 | .m(4) |
| 6172 | .n(16) |
| 6173 | .k(k) |
| 6174 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6175 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6176 | } |
| 6177 | } |
| 6178 | |
| 6179 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, k_div_16_subtile) { |
| 6180 | TEST_REQUIRES_ARM_NEON; |
| 6181 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6182 | for (uint32_t n = 1; n <= 16; n++) { |
| 6183 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6184 | GemmMicrokernelTester() |
| 6185 | .mr(4) |
| 6186 | .nr(16) |
| 6187 | .kr(4) |
| 6188 | .sr(2) |
| 6189 | .m(m) |
| 6190 | .n(n) |
| 6191 | .k(k) |
| 6192 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6193 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6194 | } |
| 6195 | } |
| 6196 | } |
| 6197 | } |
| 6198 | |
| 6199 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16) { |
| 6200 | TEST_REQUIRES_ARM_NEON; |
| 6201 | for (uint32_t n = 17; n < 32; n++) { |
| 6202 | for (size_t k = 1; k <= 80; k += 17) { |
| 6203 | GemmMicrokernelTester() |
| 6204 | .mr(4) |
| 6205 | .nr(16) |
| 6206 | .kr(4) |
| 6207 | .sr(2) |
| 6208 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6209 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6210 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6211 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6212 | } |
| 6213 | } |
| 6214 | } |
| 6215 | |
| 6216 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16_strided_cn) { |
| 6217 | TEST_REQUIRES_ARM_NEON; |
| 6218 | for (uint32_t n = 17; n < 32; n++) { |
| 6219 | for (size_t k = 1; k <= 80; k += 17) { |
| 6220 | GemmMicrokernelTester() |
| 6221 | .mr(4) |
| 6222 | .nr(16) |
| 6223 | .kr(4) |
| 6224 | .sr(2) |
| 6225 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6226 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6227 | .k(k) |
| 6228 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6229 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6230 | } |
| 6231 | } |
| 6232 | } |
| 6233 | |
| 6234 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16_strided_a) { |
| 6235 | TEST_REQUIRES_ARM_NEON; |
| 6236 | for (uint32_t n = 17; n < 32; n++) { |
| 6237 | for (size_t k = 1; k <= 80; k += 17) { |
| 6238 | GemmMicrokernelTester() |
| 6239 | .mr(4) |
| 6240 | .nr(16) |
| 6241 | .kr(4) |
| 6242 | .sr(2) |
| 6243 | .m(4) |
| 6244 | .n(n) |
| 6245 | .k(k) |
| 6246 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6247 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6248 | } |
| 6249 | } |
| 6250 | } |
| 6251 | |
| 6252 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_gt_16_subtile) { |
| 6253 | TEST_REQUIRES_ARM_NEON; |
| 6254 | for (uint32_t n = 17; n < 32; n++) { |
| 6255 | for (size_t k = 1; k <= 80; k += 17) { |
| 6256 | for (uint32_t m = 1; m <= 4; m++) { |
| 6257 | GemmMicrokernelTester() |
| 6258 | .mr(4) |
| 6259 | .nr(16) |
| 6260 | .kr(4) |
| 6261 | .sr(2) |
| 6262 | .m(m) |
| 6263 | .n(n) |
| 6264 | .k(k) |
| 6265 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6266 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6267 | } |
| 6268 | } |
| 6269 | } |
| 6270 | } |
| 6271 | |
| 6272 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16) { |
| 6273 | TEST_REQUIRES_ARM_NEON; |
| 6274 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 6275 | for (size_t k = 1; k <= 80; k += 17) { |
| 6276 | GemmMicrokernelTester() |
| 6277 | .mr(4) |
| 6278 | .nr(16) |
| 6279 | .kr(4) |
| 6280 | .sr(2) |
| 6281 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6282 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6283 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6284 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6285 | } |
| 6286 | } |
| 6287 | } |
| 6288 | |
| 6289 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16_strided_cn) { |
| 6290 | TEST_REQUIRES_ARM_NEON; |
| 6291 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 6292 | for (size_t k = 1; k <= 80; k += 17) { |
| 6293 | GemmMicrokernelTester() |
| 6294 | .mr(4) |
| 6295 | .nr(16) |
| 6296 | .kr(4) |
| 6297 | .sr(2) |
| 6298 | .m(4) |
| 6299 | .n(n) |
| 6300 | .k(k) |
| 6301 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6302 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6303 | } |
| 6304 | } |
| 6305 | } |
| 6306 | |
| 6307 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16_strided_a) { |
| 6308 | TEST_REQUIRES_ARM_NEON; |
| 6309 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 6310 | for (size_t k = 1; k <= 80; k += 17) { |
| 6311 | GemmMicrokernelTester() |
| 6312 | .mr(4) |
| 6313 | .nr(16) |
| 6314 | .kr(4) |
| 6315 | .sr(2) |
| 6316 | .m(4) |
| 6317 | .n(n) |
| 6318 | .k(k) |
| 6319 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6320 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6321 | } |
| 6322 | } |
| 6323 | } |
| 6324 | |
| 6325 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, n_div_16_subtile) { |
| 6326 | TEST_REQUIRES_ARM_NEON; |
| 6327 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 6328 | for (size_t k = 1; k <= 80; k += 17) { |
| 6329 | for (uint32_t m = 1; m <= 4; m++) { |
| 6330 | GemmMicrokernelTester() |
| 6331 | .mr(4) |
| 6332 | .nr(16) |
| 6333 | .kr(4) |
| 6334 | .sr(2) |
| 6335 | .m(m) |
| 6336 | .n(n) |
| 6337 | .k(k) |
| 6338 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6339 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6340 | } |
| 6341 | } |
| 6342 | } |
| 6343 | } |
| 6344 | |
| 6345 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, strided_cm_subtile) { |
| 6346 | TEST_REQUIRES_ARM_NEON; |
| 6347 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6348 | for (uint32_t n = 1; n <= 16; n++) { |
| 6349 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6350 | GemmMicrokernelTester() |
| 6351 | .mr(4) |
| 6352 | .nr(16) |
| 6353 | .kr(4) |
| 6354 | .sr(2) |
| 6355 | .m(m) |
| 6356 | .n(n) |
| 6357 | .k(k) |
| 6358 | .cm_stride(19) |
| 6359 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6360 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6361 | } |
| 6362 | } |
| 6363 | } |
| 6364 | } |
| 6365 | |
| 6366 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, qmin) { |
| 6367 | TEST_REQUIRES_ARM_NEON; |
| 6368 | GemmMicrokernelTester() |
| 6369 | .mr(4) |
| 6370 | .nr(16) |
| 6371 | .kr(4) |
| 6372 | .sr(2) |
| 6373 | .m(4) |
| 6374 | .n(16) |
| 6375 | .k(16) |
| 6376 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6377 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6378 | } |
| 6379 | |
| 6380 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, qmax) { |
| 6381 | TEST_REQUIRES_ARM_NEON; |
| 6382 | GemmMicrokernelTester() |
| 6383 | .mr(4) |
| 6384 | .nr(16) |
| 6385 | .kr(4) |
| 6386 | .sr(2) |
| 6387 | .m(4) |
| 6388 | .n(16) |
| 6389 | .k(16) |
| 6390 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6391 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6392 | } |
| 6393 | |
| 6394 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4S2__NEON_MLAL, strided_cm) { |
| 6395 | TEST_REQUIRES_ARM_NEON; |
| 6396 | GemmMicrokernelTester() |
| 6397 | .mr(4) |
| 6398 | .nr(16) |
| 6399 | .kr(4) |
| 6400 | .sr(2) |
| 6401 | .m(4) |
| 6402 | .n(16) |
| 6403 | .k(16) |
| 6404 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6405 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6406 | } |
| 6407 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 6408 | |
| 6409 | |
| 6410 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 6411 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8) { |
| 6412 | TEST_REQUIRES_ARM_NEON; |
| 6413 | GemmMicrokernelTester() |
| 6414 | .mr(1) |
| 6415 | .nr(8) |
| 6416 | .kr(2) |
| 6417 | .sr(4) |
| 6418 | .m(1) |
| 6419 | .n(8) |
| 6420 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6421 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6422 | } |
| 6423 | |
| 6424 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, strided_cn) { |
| 6425 | TEST_REQUIRES_ARM_NEON; |
| 6426 | GemmMicrokernelTester() |
| 6427 | .mr(1) |
| 6428 | .nr(8) |
| 6429 | .kr(2) |
| 6430 | .sr(4) |
| 6431 | .m(1) |
| 6432 | .n(8) |
| 6433 | .k(8) |
| 6434 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6435 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6436 | } |
| 6437 | |
| 6438 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8_strided_a) { |
| 6439 | TEST_REQUIRES_ARM_NEON; |
| 6440 | GemmMicrokernelTester() |
| 6441 | .mr(1) |
| 6442 | .nr(8) |
| 6443 | .kr(2) |
| 6444 | .sr(4) |
| 6445 | .m(1) |
| 6446 | .n(8) |
| 6447 | .k(8) |
| 6448 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6449 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6450 | } |
| 6451 | |
| 6452 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8_subtile) { |
| 6453 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6454 | for (uint32_t n = 1; n <= 8; n++) { |
| 6455 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6456 | GemmMicrokernelTester() |
| 6457 | .mr(1) |
| 6458 | .nr(8) |
| 6459 | .kr(2) |
| 6460 | .sr(4) |
| 6461 | .m(m) |
| 6462 | .n(n) |
| 6463 | .k(8) |
| 6464 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6465 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6466 | } |
| 6467 | } |
| 6468 | } |
| 6469 | |
| 6470 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8_subtile_m) { |
| 6471 | TEST_REQUIRES_ARM_NEON; |
| 6472 | for (uint32_t m = 1; m <= 1; m++) { |
| 6473 | GemmMicrokernelTester() |
| 6474 | .mr(1) |
| 6475 | .nr(8) |
| 6476 | .kr(2) |
| 6477 | .sr(4) |
| 6478 | .m(m) |
| 6479 | .n(8) |
| 6480 | .k(8) |
| 6481 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6482 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6483 | } |
| 6484 | } |
| 6485 | |
| 6486 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_eq_8_subtile_n) { |
| 6487 | TEST_REQUIRES_ARM_NEON; |
| 6488 | for (uint32_t n = 1; n <= 8; n++) { |
| 6489 | GemmMicrokernelTester() |
| 6490 | .mr(1) |
| 6491 | .nr(8) |
| 6492 | .kr(2) |
| 6493 | .sr(4) |
| 6494 | .m(1) |
| 6495 | .n(n) |
| 6496 | .k(8) |
| 6497 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6498 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6499 | } |
| 6500 | } |
| 6501 | |
| 6502 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_lt_8) { |
| 6503 | TEST_REQUIRES_ARM_NEON; |
| 6504 | for (size_t k = 1; k < 8; k++) { |
| 6505 | GemmMicrokernelTester() |
| 6506 | .mr(1) |
| 6507 | .nr(8) |
| 6508 | .kr(2) |
| 6509 | .sr(4) |
| 6510 | .m(1) |
| 6511 | .n(8) |
| 6512 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6513 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6514 | } |
| 6515 | } |
| 6516 | |
| 6517 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_lt_8_strided_a) { |
| 6518 | TEST_REQUIRES_ARM_NEON; |
| 6519 | for (size_t k = 1; k < 8; k++) { |
| 6520 | GemmMicrokernelTester() |
| 6521 | .mr(1) |
| 6522 | .nr(8) |
| 6523 | .kr(2) |
| 6524 | .sr(4) |
| 6525 | .m(1) |
| 6526 | .n(8) |
| 6527 | .k(k) |
| 6528 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6529 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6530 | } |
| 6531 | } |
| 6532 | |
| 6533 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_lt_8_subtile) { |
| 6534 | TEST_REQUIRES_ARM_NEON; |
| 6535 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6536 | for (uint32_t n = 1; n <= 8; n++) { |
| 6537 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6538 | GemmMicrokernelTester() |
| 6539 | .mr(1) |
| 6540 | .nr(8) |
| 6541 | .kr(2) |
| 6542 | .sr(4) |
| 6543 | .m(m) |
| 6544 | .n(n) |
| 6545 | .k(k) |
| 6546 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6547 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6548 | } |
| 6549 | } |
| 6550 | } |
| 6551 | } |
| 6552 | |
| 6553 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_gt_8) { |
| 6554 | TEST_REQUIRES_ARM_NEON; |
| 6555 | for (size_t k = 9; k < 16; k++) { |
| 6556 | GemmMicrokernelTester() |
| 6557 | .mr(1) |
| 6558 | .nr(8) |
| 6559 | .kr(2) |
| 6560 | .sr(4) |
| 6561 | .m(1) |
| 6562 | .n(8) |
| 6563 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6564 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6565 | } |
| 6566 | } |
| 6567 | |
| 6568 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_gt_8_strided_a) { |
| 6569 | TEST_REQUIRES_ARM_NEON; |
| 6570 | for (size_t k = 9; k < 16; k++) { |
| 6571 | GemmMicrokernelTester() |
| 6572 | .mr(1) |
| 6573 | .nr(8) |
| 6574 | .kr(2) |
| 6575 | .sr(4) |
| 6576 | .m(1) |
| 6577 | .n(8) |
| 6578 | .k(k) |
| 6579 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6580 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6581 | } |
| 6582 | } |
| 6583 | |
| 6584 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_gt_8_subtile) { |
| 6585 | TEST_REQUIRES_ARM_NEON; |
| 6586 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6587 | for (uint32_t n = 1; n <= 8; n++) { |
| 6588 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6589 | GemmMicrokernelTester() |
| 6590 | .mr(1) |
| 6591 | .nr(8) |
| 6592 | .kr(2) |
| 6593 | .sr(4) |
| 6594 | .m(m) |
| 6595 | .n(n) |
| 6596 | .k(k) |
| 6597 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6598 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6599 | } |
| 6600 | } |
| 6601 | } |
| 6602 | } |
| 6603 | |
| 6604 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_div_8) { |
| 6605 | TEST_REQUIRES_ARM_NEON; |
| 6606 | for (size_t k = 16; k <= 80; k += 8) { |
| 6607 | GemmMicrokernelTester() |
| 6608 | .mr(1) |
| 6609 | .nr(8) |
| 6610 | .kr(2) |
| 6611 | .sr(4) |
| 6612 | .m(1) |
| 6613 | .n(8) |
| 6614 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6615 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6616 | } |
| 6617 | } |
| 6618 | |
| 6619 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_div_8_strided_a) { |
| 6620 | TEST_REQUIRES_ARM_NEON; |
| 6621 | for (size_t k = 16; k <= 80; k += 8) { |
| 6622 | GemmMicrokernelTester() |
| 6623 | .mr(1) |
| 6624 | .nr(8) |
| 6625 | .kr(2) |
| 6626 | .sr(4) |
| 6627 | .m(1) |
| 6628 | .n(8) |
| 6629 | .k(k) |
| 6630 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6631 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6632 | } |
| 6633 | } |
| 6634 | |
| 6635 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, k_div_8_subtile) { |
| 6636 | TEST_REQUIRES_ARM_NEON; |
| 6637 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6638 | for (uint32_t n = 1; n <= 8; n++) { |
| 6639 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6640 | GemmMicrokernelTester() |
| 6641 | .mr(1) |
| 6642 | .nr(8) |
| 6643 | .kr(2) |
| 6644 | .sr(4) |
| 6645 | .m(m) |
| 6646 | .n(n) |
| 6647 | .k(k) |
| 6648 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6649 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6650 | } |
| 6651 | } |
| 6652 | } |
| 6653 | } |
| 6654 | |
| 6655 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_gt_8) { |
| 6656 | TEST_REQUIRES_ARM_NEON; |
| 6657 | for (uint32_t n = 9; n < 16; n++) { |
| 6658 | for (size_t k = 1; k <= 40; k += 9) { |
| 6659 | GemmMicrokernelTester() |
| 6660 | .mr(1) |
| 6661 | .nr(8) |
| 6662 | .kr(2) |
| 6663 | .sr(4) |
| 6664 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6665 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6666 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6667 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6668 | } |
| 6669 | } |
| 6670 | } |
| 6671 | |
| 6672 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_gt_8_strided_cn) { |
| 6673 | TEST_REQUIRES_ARM_NEON; |
| 6674 | for (uint32_t n = 9; n < 16; n++) { |
| 6675 | for (size_t k = 1; k <= 40; k += 9) { |
| 6676 | GemmMicrokernelTester() |
| 6677 | .mr(1) |
| 6678 | .nr(8) |
| 6679 | .kr(2) |
| 6680 | .sr(4) |
| 6681 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6682 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6683 | .k(k) |
| 6684 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6685 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6686 | } |
| 6687 | } |
| 6688 | } |
| 6689 | |
| 6690 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_gt_8_strided_a) { |
| 6691 | TEST_REQUIRES_ARM_NEON; |
| 6692 | for (uint32_t n = 9; n < 16; n++) { |
| 6693 | for (size_t k = 1; k <= 40; k += 9) { |
| 6694 | GemmMicrokernelTester() |
| 6695 | .mr(1) |
| 6696 | .nr(8) |
| 6697 | .kr(2) |
| 6698 | .sr(4) |
| 6699 | .m(1) |
| 6700 | .n(n) |
| 6701 | .k(k) |
| 6702 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6703 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6704 | } |
| 6705 | } |
| 6706 | } |
| 6707 | |
| 6708 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_gt_8_subtile) { |
| 6709 | TEST_REQUIRES_ARM_NEON; |
| 6710 | for (uint32_t n = 9; n < 16; n++) { |
| 6711 | for (size_t k = 1; k <= 40; k += 9) { |
| 6712 | for (uint32_t m = 1; m <= 1; m++) { |
| 6713 | GemmMicrokernelTester() |
| 6714 | .mr(1) |
| 6715 | .nr(8) |
| 6716 | .kr(2) |
| 6717 | .sr(4) |
| 6718 | .m(m) |
| 6719 | .n(n) |
| 6720 | .k(k) |
| 6721 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6722 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6723 | } |
| 6724 | } |
| 6725 | } |
| 6726 | } |
| 6727 | |
| 6728 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_div_8) { |
| 6729 | TEST_REQUIRES_ARM_NEON; |
| 6730 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6731 | for (size_t k = 1; k <= 40; k += 9) { |
| 6732 | GemmMicrokernelTester() |
| 6733 | .mr(1) |
| 6734 | .nr(8) |
| 6735 | .kr(2) |
| 6736 | .sr(4) |
| 6737 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 6738 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6739 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6740 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6741 | } |
| 6742 | } |
| 6743 | } |
| 6744 | |
| 6745 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_div_8_strided_cn) { |
| 6746 | TEST_REQUIRES_ARM_NEON; |
| 6747 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6748 | for (size_t k = 1; k <= 40; k += 9) { |
| 6749 | GemmMicrokernelTester() |
| 6750 | .mr(1) |
| 6751 | .nr(8) |
| 6752 | .kr(2) |
| 6753 | .sr(4) |
| 6754 | .m(1) |
| 6755 | .n(n) |
| 6756 | .k(k) |
| 6757 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6758 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6759 | } |
| 6760 | } |
| 6761 | } |
| 6762 | |
| 6763 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_div_8_strided_a) { |
| 6764 | TEST_REQUIRES_ARM_NEON; |
| 6765 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6766 | for (size_t k = 1; k <= 40; k += 9) { |
| 6767 | GemmMicrokernelTester() |
| 6768 | .mr(1) |
| 6769 | .nr(8) |
| 6770 | .kr(2) |
| 6771 | .sr(4) |
| 6772 | .m(1) |
| 6773 | .n(n) |
| 6774 | .k(k) |
| 6775 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6776 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6777 | } |
| 6778 | } |
| 6779 | } |
| 6780 | |
| 6781 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, n_div_8_subtile) { |
| 6782 | TEST_REQUIRES_ARM_NEON; |
| 6783 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 6784 | for (size_t k = 1; k <= 40; k += 9) { |
| 6785 | for (uint32_t m = 1; m <= 1; m++) { |
| 6786 | GemmMicrokernelTester() |
| 6787 | .mr(1) |
| 6788 | .nr(8) |
| 6789 | .kr(2) |
| 6790 | .sr(4) |
| 6791 | .m(m) |
| 6792 | .n(n) |
| 6793 | .k(k) |
| 6794 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6795 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6796 | } |
| 6797 | } |
| 6798 | } |
| 6799 | } |
| 6800 | |
| 6801 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, strided_cm_subtile) { |
| 6802 | TEST_REQUIRES_ARM_NEON; |
| 6803 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6804 | for (uint32_t n = 1; n <= 8; n++) { |
| 6805 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6806 | GemmMicrokernelTester() |
| 6807 | .mr(1) |
| 6808 | .nr(8) |
| 6809 | .kr(2) |
| 6810 | .sr(4) |
| 6811 | .m(m) |
| 6812 | .n(n) |
| 6813 | .k(k) |
| 6814 | .cm_stride(11) |
| 6815 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6816 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6817 | } |
| 6818 | } |
| 6819 | } |
| 6820 | } |
| 6821 | |
| 6822 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, qmin) { |
| 6823 | TEST_REQUIRES_ARM_NEON; |
| 6824 | GemmMicrokernelTester() |
| 6825 | .mr(1) |
| 6826 | .nr(8) |
| 6827 | .kr(2) |
| 6828 | .sr(4) |
| 6829 | .m(1) |
| 6830 | .n(8) |
| 6831 | .k(8) |
| 6832 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6833 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6834 | } |
| 6835 | |
| 6836 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, qmax) { |
| 6837 | TEST_REQUIRES_ARM_NEON; |
| 6838 | GemmMicrokernelTester() |
| 6839 | .mr(1) |
| 6840 | .nr(8) |
| 6841 | .kr(2) |
| 6842 | .sr(4) |
| 6843 | .m(1) |
| 6844 | .n(8) |
| 6845 | .k(8) |
| 6846 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6847 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6848 | } |
| 6849 | |
| 6850 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C2S4__NEON_MULL, strided_cm) { |
| 6851 | TEST_REQUIRES_ARM_NEON; |
| 6852 | GemmMicrokernelTester() |
| 6853 | .mr(1) |
| 6854 | .nr(8) |
| 6855 | .kr(2) |
| 6856 | .sr(4) |
| 6857 | .m(1) |
| 6858 | .n(8) |
| 6859 | .k(8) |
| 6860 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6861 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6862 | } |
| 6863 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 6864 | |
| 6865 | |
| 6866 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 6867 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8) { |
| 6868 | TEST_REQUIRES_ARM_NEON; |
| 6869 | GemmMicrokernelTester() |
| 6870 | .mr(4) |
| 6871 | .nr(8) |
| 6872 | .kr(2) |
| 6873 | .sr(4) |
| 6874 | .m(4) |
| 6875 | .n(8) |
| 6876 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6877 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6878 | } |
| 6879 | |
| 6880 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, strided_cn) { |
| 6881 | TEST_REQUIRES_ARM_NEON; |
| 6882 | GemmMicrokernelTester() |
| 6883 | .mr(4) |
| 6884 | .nr(8) |
| 6885 | .kr(2) |
| 6886 | .sr(4) |
| 6887 | .m(4) |
| 6888 | .n(8) |
| 6889 | .k(8) |
| 6890 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6891 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6892 | } |
| 6893 | |
| 6894 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8_strided_a) { |
| 6895 | TEST_REQUIRES_ARM_NEON; |
| 6896 | GemmMicrokernelTester() |
| 6897 | .mr(4) |
| 6898 | .nr(8) |
| 6899 | .kr(2) |
| 6900 | .sr(4) |
| 6901 | .m(4) |
| 6902 | .n(8) |
| 6903 | .k(8) |
| 6904 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6905 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6906 | } |
| 6907 | |
| 6908 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8_subtile) { |
| 6909 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6910 | for (uint32_t n = 1; n <= 8; n++) { |
| 6911 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6912 | GemmMicrokernelTester() |
| 6913 | .mr(4) |
| 6914 | .nr(8) |
| 6915 | .kr(2) |
| 6916 | .sr(4) |
| 6917 | .m(m) |
| 6918 | .n(n) |
| 6919 | .k(8) |
| 6920 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6921 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6922 | } |
| 6923 | } |
| 6924 | } |
| 6925 | |
| 6926 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8_subtile_m) { |
| 6927 | TEST_REQUIRES_ARM_NEON; |
| 6928 | for (uint32_t m = 1; m <= 4; m++) { |
| 6929 | GemmMicrokernelTester() |
| 6930 | .mr(4) |
| 6931 | .nr(8) |
| 6932 | .kr(2) |
| 6933 | .sr(4) |
| 6934 | .m(m) |
| 6935 | .n(8) |
| 6936 | .k(8) |
| 6937 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6938 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6939 | } |
| 6940 | } |
| 6941 | |
| 6942 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_eq_8_subtile_n) { |
| 6943 | TEST_REQUIRES_ARM_NEON; |
| 6944 | for (uint32_t n = 1; n <= 8; n++) { |
| 6945 | GemmMicrokernelTester() |
| 6946 | .mr(4) |
| 6947 | .nr(8) |
| 6948 | .kr(2) |
| 6949 | .sr(4) |
| 6950 | .m(4) |
| 6951 | .n(n) |
| 6952 | .k(8) |
| 6953 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6954 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6955 | } |
| 6956 | } |
| 6957 | |
| 6958 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_lt_8) { |
| 6959 | TEST_REQUIRES_ARM_NEON; |
| 6960 | for (size_t k = 1; k < 8; k++) { |
| 6961 | GemmMicrokernelTester() |
| 6962 | .mr(4) |
| 6963 | .nr(8) |
| 6964 | .kr(2) |
| 6965 | .sr(4) |
| 6966 | .m(4) |
| 6967 | .n(8) |
| 6968 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6969 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6970 | } |
| 6971 | } |
| 6972 | |
| 6973 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_lt_8_strided_a) { |
| 6974 | TEST_REQUIRES_ARM_NEON; |
| 6975 | for (size_t k = 1; k < 8; k++) { |
| 6976 | GemmMicrokernelTester() |
| 6977 | .mr(4) |
| 6978 | .nr(8) |
| 6979 | .kr(2) |
| 6980 | .sr(4) |
| 6981 | .m(4) |
| 6982 | .n(8) |
| 6983 | .k(k) |
| 6984 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 6985 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6986 | } |
| 6987 | } |
| 6988 | |
| 6989 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_lt_8_subtile) { |
| 6990 | TEST_REQUIRES_ARM_NEON; |
| 6991 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 6992 | for (uint32_t n = 1; n <= 8; n++) { |
| 6993 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 6994 | GemmMicrokernelTester() |
| 6995 | .mr(4) |
| 6996 | .nr(8) |
| 6997 | .kr(2) |
| 6998 | .sr(4) |
| 6999 | .m(m) |
| 7000 | .n(n) |
| 7001 | .k(k) |
| 7002 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7003 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7004 | } |
| 7005 | } |
| 7006 | } |
| 7007 | } |
| 7008 | |
| 7009 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_gt_8) { |
| 7010 | TEST_REQUIRES_ARM_NEON; |
| 7011 | for (size_t k = 9; k < 16; k++) { |
| 7012 | GemmMicrokernelTester() |
| 7013 | .mr(4) |
| 7014 | .nr(8) |
| 7015 | .kr(2) |
| 7016 | .sr(4) |
| 7017 | .m(4) |
| 7018 | .n(8) |
| 7019 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7020 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7021 | } |
| 7022 | } |
| 7023 | |
| 7024 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_gt_8_strided_a) { |
| 7025 | TEST_REQUIRES_ARM_NEON; |
| 7026 | for (size_t k = 9; k < 16; k++) { |
| 7027 | GemmMicrokernelTester() |
| 7028 | .mr(4) |
| 7029 | .nr(8) |
| 7030 | .kr(2) |
| 7031 | .sr(4) |
| 7032 | .m(4) |
| 7033 | .n(8) |
| 7034 | .k(k) |
| 7035 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7036 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7037 | } |
| 7038 | } |
| 7039 | |
| 7040 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_gt_8_subtile) { |
| 7041 | TEST_REQUIRES_ARM_NEON; |
| 7042 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7043 | for (uint32_t n = 1; n <= 8; n++) { |
| 7044 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7045 | GemmMicrokernelTester() |
| 7046 | .mr(4) |
| 7047 | .nr(8) |
| 7048 | .kr(2) |
| 7049 | .sr(4) |
| 7050 | .m(m) |
| 7051 | .n(n) |
| 7052 | .k(k) |
| 7053 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7054 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7055 | } |
| 7056 | } |
| 7057 | } |
| 7058 | } |
| 7059 | |
| 7060 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_div_8) { |
| 7061 | TEST_REQUIRES_ARM_NEON; |
| 7062 | for (size_t k = 16; k <= 80; k += 8) { |
| 7063 | GemmMicrokernelTester() |
| 7064 | .mr(4) |
| 7065 | .nr(8) |
| 7066 | .kr(2) |
| 7067 | .sr(4) |
| 7068 | .m(4) |
| 7069 | .n(8) |
| 7070 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7071 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7072 | } |
| 7073 | } |
| 7074 | |
| 7075 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_div_8_strided_a) { |
| 7076 | TEST_REQUIRES_ARM_NEON; |
| 7077 | for (size_t k = 16; k <= 80; k += 8) { |
| 7078 | GemmMicrokernelTester() |
| 7079 | .mr(4) |
| 7080 | .nr(8) |
| 7081 | .kr(2) |
| 7082 | .sr(4) |
| 7083 | .m(4) |
| 7084 | .n(8) |
| 7085 | .k(k) |
| 7086 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7087 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7088 | } |
| 7089 | } |
| 7090 | |
| 7091 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, k_div_8_subtile) { |
| 7092 | TEST_REQUIRES_ARM_NEON; |
| 7093 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7094 | for (uint32_t n = 1; n <= 8; n++) { |
| 7095 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7096 | GemmMicrokernelTester() |
| 7097 | .mr(4) |
| 7098 | .nr(8) |
| 7099 | .kr(2) |
| 7100 | .sr(4) |
| 7101 | .m(m) |
| 7102 | .n(n) |
| 7103 | .k(k) |
| 7104 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7105 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7106 | } |
| 7107 | } |
| 7108 | } |
| 7109 | } |
| 7110 | |
| 7111 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8) { |
| 7112 | TEST_REQUIRES_ARM_NEON; |
| 7113 | for (uint32_t n = 9; n < 16; n++) { |
| 7114 | for (size_t k = 1; k <= 40; k += 9) { |
| 7115 | GemmMicrokernelTester() |
| 7116 | .mr(4) |
| 7117 | .nr(8) |
| 7118 | .kr(2) |
| 7119 | .sr(4) |
| 7120 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7121 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7122 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7123 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7124 | } |
| 7125 | } |
| 7126 | } |
| 7127 | |
| 7128 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8_strided_cn) { |
| 7129 | TEST_REQUIRES_ARM_NEON; |
| 7130 | for (uint32_t n = 9; n < 16; n++) { |
| 7131 | for (size_t k = 1; k <= 40; k += 9) { |
| 7132 | GemmMicrokernelTester() |
| 7133 | .mr(4) |
| 7134 | .nr(8) |
| 7135 | .kr(2) |
| 7136 | .sr(4) |
| 7137 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7138 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7139 | .k(k) |
| 7140 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7141 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7142 | } |
| 7143 | } |
| 7144 | } |
| 7145 | |
| 7146 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8_strided_a) { |
| 7147 | TEST_REQUIRES_ARM_NEON; |
| 7148 | for (uint32_t n = 9; n < 16; n++) { |
| 7149 | for (size_t k = 1; k <= 40; k += 9) { |
| 7150 | GemmMicrokernelTester() |
| 7151 | .mr(4) |
| 7152 | .nr(8) |
| 7153 | .kr(2) |
| 7154 | .sr(4) |
| 7155 | .m(4) |
| 7156 | .n(n) |
| 7157 | .k(k) |
| 7158 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7159 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7160 | } |
| 7161 | } |
| 7162 | } |
| 7163 | |
| 7164 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_gt_8_subtile) { |
| 7165 | TEST_REQUIRES_ARM_NEON; |
| 7166 | for (uint32_t n = 9; n < 16; n++) { |
| 7167 | for (size_t k = 1; k <= 40; k += 9) { |
| 7168 | for (uint32_t m = 1; m <= 4; m++) { |
| 7169 | GemmMicrokernelTester() |
| 7170 | .mr(4) |
| 7171 | .nr(8) |
| 7172 | .kr(2) |
| 7173 | .sr(4) |
| 7174 | .m(m) |
| 7175 | .n(n) |
| 7176 | .k(k) |
| 7177 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7178 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7179 | } |
| 7180 | } |
| 7181 | } |
| 7182 | } |
| 7183 | |
| 7184 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8) { |
| 7185 | TEST_REQUIRES_ARM_NEON; |
| 7186 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7187 | for (size_t k = 1; k <= 40; k += 9) { |
| 7188 | GemmMicrokernelTester() |
| 7189 | .mr(4) |
| 7190 | .nr(8) |
| 7191 | .kr(2) |
| 7192 | .sr(4) |
| 7193 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7194 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7195 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7196 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7197 | } |
| 7198 | } |
| 7199 | } |
| 7200 | |
| 7201 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8_strided_cn) { |
| 7202 | TEST_REQUIRES_ARM_NEON; |
| 7203 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7204 | for (size_t k = 1; k <= 40; k += 9) { |
| 7205 | GemmMicrokernelTester() |
| 7206 | .mr(4) |
| 7207 | .nr(8) |
| 7208 | .kr(2) |
| 7209 | .sr(4) |
| 7210 | .m(4) |
| 7211 | .n(n) |
| 7212 | .k(k) |
| 7213 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7214 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7215 | } |
| 7216 | } |
| 7217 | } |
| 7218 | |
| 7219 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8_strided_a) { |
| 7220 | TEST_REQUIRES_ARM_NEON; |
| 7221 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7222 | for (size_t k = 1; k <= 40; k += 9) { |
| 7223 | GemmMicrokernelTester() |
| 7224 | .mr(4) |
| 7225 | .nr(8) |
| 7226 | .kr(2) |
| 7227 | .sr(4) |
| 7228 | .m(4) |
| 7229 | .n(n) |
| 7230 | .k(k) |
| 7231 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7232 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7233 | } |
| 7234 | } |
| 7235 | } |
| 7236 | |
| 7237 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, n_div_8_subtile) { |
| 7238 | TEST_REQUIRES_ARM_NEON; |
| 7239 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 7240 | for (size_t k = 1; k <= 40; k += 9) { |
| 7241 | for (uint32_t m = 1; m <= 4; m++) { |
| 7242 | GemmMicrokernelTester() |
| 7243 | .mr(4) |
| 7244 | .nr(8) |
| 7245 | .kr(2) |
| 7246 | .sr(4) |
| 7247 | .m(m) |
| 7248 | .n(n) |
| 7249 | .k(k) |
| 7250 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7251 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7252 | } |
| 7253 | } |
| 7254 | } |
| 7255 | } |
| 7256 | |
| 7257 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, strided_cm_subtile) { |
| 7258 | TEST_REQUIRES_ARM_NEON; |
| 7259 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7260 | for (uint32_t n = 1; n <= 8; n++) { |
| 7261 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7262 | GemmMicrokernelTester() |
| 7263 | .mr(4) |
| 7264 | .nr(8) |
| 7265 | .kr(2) |
| 7266 | .sr(4) |
| 7267 | .m(m) |
| 7268 | .n(n) |
| 7269 | .k(k) |
| 7270 | .cm_stride(11) |
| 7271 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7272 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7273 | } |
| 7274 | } |
| 7275 | } |
| 7276 | } |
| 7277 | |
| 7278 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, qmin) { |
| 7279 | TEST_REQUIRES_ARM_NEON; |
| 7280 | GemmMicrokernelTester() |
| 7281 | .mr(4) |
| 7282 | .nr(8) |
| 7283 | .kr(2) |
| 7284 | .sr(4) |
| 7285 | .m(4) |
| 7286 | .n(8) |
| 7287 | .k(8) |
| 7288 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7289 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7290 | } |
| 7291 | |
| 7292 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, qmax) { |
| 7293 | TEST_REQUIRES_ARM_NEON; |
| 7294 | GemmMicrokernelTester() |
| 7295 | .mr(4) |
| 7296 | .nr(8) |
| 7297 | .kr(2) |
| 7298 | .sr(4) |
| 7299 | .m(4) |
| 7300 | .n(8) |
| 7301 | .k(8) |
| 7302 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7303 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7304 | } |
| 7305 | |
| 7306 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2S4__NEON_MULL, strided_cm) { |
| 7307 | TEST_REQUIRES_ARM_NEON; |
| 7308 | GemmMicrokernelTester() |
| 7309 | .mr(4) |
| 7310 | .nr(8) |
| 7311 | .kr(2) |
| 7312 | .sr(4) |
| 7313 | .m(4) |
| 7314 | .n(8) |
| 7315 | .k(8) |
| 7316 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7317 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7318 | } |
| 7319 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 7320 | |
| 7321 | |
| 7322 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 7323 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_eq_8) { |
| 7324 | TEST_REQUIRES_ARM_NEON; |
| 7325 | GemmMicrokernelTester() |
| 7326 | .mr(3) |
| 7327 | .nr(16) |
| 7328 | .kr(2) |
| 7329 | .sr(4) |
| 7330 | .m(3) |
| 7331 | .n(16) |
| 7332 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7333 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7334 | } |
| 7335 | |
| 7336 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, strided_cn) { |
| 7337 | TEST_REQUIRES_ARM_NEON; |
| 7338 | GemmMicrokernelTester() |
| 7339 | .mr(3) |
| 7340 | .nr(16) |
| 7341 | .kr(2) |
| 7342 | .sr(4) |
| 7343 | .m(3) |
| 7344 | .n(16) |
| 7345 | .k(8) |
| 7346 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7347 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7348 | } |
| 7349 | |
| 7350 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_eq_8_strided_a) { |
| 7351 | TEST_REQUIRES_ARM_NEON; |
| 7352 | GemmMicrokernelTester() |
| 7353 | .mr(3) |
| 7354 | .nr(16) |
| 7355 | .kr(2) |
| 7356 | .sr(4) |
| 7357 | .m(3) |
| 7358 | .n(16) |
| 7359 | .k(8) |
| 7360 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7361 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7362 | } |
| 7363 | |
| 7364 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_eq_8_subtile) { |
| 7365 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7366 | for (uint32_t n = 1; n <= 16; n++) { |
| 7367 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7368 | GemmMicrokernelTester() |
| 7369 | .mr(3) |
| 7370 | .nr(16) |
| 7371 | .kr(2) |
| 7372 | .sr(4) |
| 7373 | .m(m) |
| 7374 | .n(n) |
| 7375 | .k(8) |
| 7376 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7377 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7378 | } |
| 7379 | } |
| 7380 | } |
| 7381 | |
| 7382 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_eq_8_subtile_m) { |
| 7383 | TEST_REQUIRES_ARM_NEON; |
| 7384 | for (uint32_t m = 1; m <= 3; m++) { |
| 7385 | GemmMicrokernelTester() |
| 7386 | .mr(3) |
| 7387 | .nr(16) |
| 7388 | .kr(2) |
| 7389 | .sr(4) |
| 7390 | .m(m) |
| 7391 | .n(16) |
| 7392 | .k(8) |
| 7393 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7394 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7395 | } |
| 7396 | } |
| 7397 | |
| 7398 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_eq_8_subtile_n) { |
| 7399 | TEST_REQUIRES_ARM_NEON; |
| 7400 | for (uint32_t n = 1; n <= 16; n++) { |
| 7401 | GemmMicrokernelTester() |
| 7402 | .mr(3) |
| 7403 | .nr(16) |
| 7404 | .kr(2) |
| 7405 | .sr(4) |
| 7406 | .m(3) |
| 7407 | .n(n) |
| 7408 | .k(8) |
| 7409 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7410 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7411 | } |
| 7412 | } |
| 7413 | |
| 7414 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_lt_8) { |
| 7415 | TEST_REQUIRES_ARM_NEON; |
| 7416 | for (size_t k = 1; k < 8; k++) { |
| 7417 | GemmMicrokernelTester() |
| 7418 | .mr(3) |
| 7419 | .nr(16) |
| 7420 | .kr(2) |
| 7421 | .sr(4) |
| 7422 | .m(3) |
| 7423 | .n(16) |
| 7424 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7425 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7426 | } |
| 7427 | } |
| 7428 | |
| 7429 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_lt_8_strided_a) { |
| 7430 | TEST_REQUIRES_ARM_NEON; |
| 7431 | for (size_t k = 1; k < 8; k++) { |
| 7432 | GemmMicrokernelTester() |
| 7433 | .mr(3) |
| 7434 | .nr(16) |
| 7435 | .kr(2) |
| 7436 | .sr(4) |
| 7437 | .m(3) |
| 7438 | .n(16) |
| 7439 | .k(k) |
| 7440 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7441 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7442 | } |
| 7443 | } |
| 7444 | |
| 7445 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_lt_8_subtile) { |
| 7446 | TEST_REQUIRES_ARM_NEON; |
| 7447 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7448 | for (uint32_t n = 1; n <= 16; n++) { |
| 7449 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7450 | GemmMicrokernelTester() |
| 7451 | .mr(3) |
| 7452 | .nr(16) |
| 7453 | .kr(2) |
| 7454 | .sr(4) |
| 7455 | .m(m) |
| 7456 | .n(n) |
| 7457 | .k(k) |
| 7458 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7459 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7460 | } |
| 7461 | } |
| 7462 | } |
| 7463 | } |
| 7464 | |
| 7465 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_gt_8) { |
| 7466 | TEST_REQUIRES_ARM_NEON; |
| 7467 | for (size_t k = 9; k < 16; k++) { |
| 7468 | GemmMicrokernelTester() |
| 7469 | .mr(3) |
| 7470 | .nr(16) |
| 7471 | .kr(2) |
| 7472 | .sr(4) |
| 7473 | .m(3) |
| 7474 | .n(16) |
| 7475 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7476 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7477 | } |
| 7478 | } |
| 7479 | |
| 7480 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_gt_8_strided_a) { |
| 7481 | TEST_REQUIRES_ARM_NEON; |
| 7482 | for (size_t k = 9; k < 16; k++) { |
| 7483 | GemmMicrokernelTester() |
| 7484 | .mr(3) |
| 7485 | .nr(16) |
| 7486 | .kr(2) |
| 7487 | .sr(4) |
| 7488 | .m(3) |
| 7489 | .n(16) |
| 7490 | .k(k) |
| 7491 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7492 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7493 | } |
| 7494 | } |
| 7495 | |
| 7496 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_gt_8_subtile) { |
| 7497 | TEST_REQUIRES_ARM_NEON; |
| 7498 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7499 | for (uint32_t n = 1; n <= 16; n++) { |
| 7500 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7501 | GemmMicrokernelTester() |
| 7502 | .mr(3) |
| 7503 | .nr(16) |
| 7504 | .kr(2) |
| 7505 | .sr(4) |
| 7506 | .m(m) |
| 7507 | .n(n) |
| 7508 | .k(k) |
| 7509 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7510 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7511 | } |
| 7512 | } |
| 7513 | } |
| 7514 | } |
| 7515 | |
| 7516 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_div_8) { |
| 7517 | TEST_REQUIRES_ARM_NEON; |
| 7518 | for (size_t k = 16; k <= 80; k += 8) { |
| 7519 | GemmMicrokernelTester() |
| 7520 | .mr(3) |
| 7521 | .nr(16) |
| 7522 | .kr(2) |
| 7523 | .sr(4) |
| 7524 | .m(3) |
| 7525 | .n(16) |
| 7526 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7527 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7528 | } |
| 7529 | } |
| 7530 | |
| 7531 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_div_8_strided_a) { |
| 7532 | TEST_REQUIRES_ARM_NEON; |
| 7533 | for (size_t k = 16; k <= 80; k += 8) { |
| 7534 | GemmMicrokernelTester() |
| 7535 | .mr(3) |
| 7536 | .nr(16) |
| 7537 | .kr(2) |
| 7538 | .sr(4) |
| 7539 | .m(3) |
| 7540 | .n(16) |
| 7541 | .k(k) |
| 7542 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7543 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7544 | } |
| 7545 | } |
| 7546 | |
| 7547 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, k_div_8_subtile) { |
| 7548 | TEST_REQUIRES_ARM_NEON; |
| 7549 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7550 | for (uint32_t n = 1; n <= 16; n++) { |
| 7551 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7552 | GemmMicrokernelTester() |
| 7553 | .mr(3) |
| 7554 | .nr(16) |
| 7555 | .kr(2) |
| 7556 | .sr(4) |
| 7557 | .m(m) |
| 7558 | .n(n) |
| 7559 | .k(k) |
| 7560 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7561 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7562 | } |
| 7563 | } |
| 7564 | } |
| 7565 | } |
| 7566 | |
| 7567 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_gt_16) { |
| 7568 | TEST_REQUIRES_ARM_NEON; |
| 7569 | for (uint32_t n = 17; n < 32; n++) { |
| 7570 | for (size_t k = 1; k <= 40; k += 9) { |
| 7571 | GemmMicrokernelTester() |
| 7572 | .mr(3) |
| 7573 | .nr(16) |
| 7574 | .kr(2) |
| 7575 | .sr(4) |
| 7576 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7577 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7578 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7579 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7580 | } |
| 7581 | } |
| 7582 | } |
| 7583 | |
| 7584 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_gt_16_strided_cn) { |
| 7585 | TEST_REQUIRES_ARM_NEON; |
| 7586 | for (uint32_t n = 17; n < 32; n++) { |
| 7587 | for (size_t k = 1; k <= 40; k += 9) { |
| 7588 | GemmMicrokernelTester() |
| 7589 | .mr(3) |
| 7590 | .nr(16) |
| 7591 | .kr(2) |
| 7592 | .sr(4) |
| 7593 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7594 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7595 | .k(k) |
| 7596 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7597 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7598 | } |
| 7599 | } |
| 7600 | } |
| 7601 | |
| 7602 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_gt_16_strided_a) { |
| 7603 | TEST_REQUIRES_ARM_NEON; |
| 7604 | for (uint32_t n = 17; n < 32; n++) { |
| 7605 | for (size_t k = 1; k <= 40; k += 9) { |
| 7606 | GemmMicrokernelTester() |
| 7607 | .mr(3) |
| 7608 | .nr(16) |
| 7609 | .kr(2) |
| 7610 | .sr(4) |
| 7611 | .m(3) |
| 7612 | .n(n) |
| 7613 | .k(k) |
| 7614 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7615 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7616 | } |
| 7617 | } |
| 7618 | } |
| 7619 | |
| 7620 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_gt_16_subtile) { |
| 7621 | TEST_REQUIRES_ARM_NEON; |
| 7622 | for (uint32_t n = 17; n < 32; n++) { |
| 7623 | for (size_t k = 1; k <= 40; k += 9) { |
| 7624 | for (uint32_t m = 1; m <= 3; m++) { |
| 7625 | GemmMicrokernelTester() |
| 7626 | .mr(3) |
| 7627 | .nr(16) |
| 7628 | .kr(2) |
| 7629 | .sr(4) |
| 7630 | .m(m) |
| 7631 | .n(n) |
| 7632 | .k(k) |
| 7633 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7634 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7635 | } |
| 7636 | } |
| 7637 | } |
| 7638 | } |
| 7639 | |
| 7640 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_div_16) { |
| 7641 | TEST_REQUIRES_ARM_NEON; |
| 7642 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 7643 | for (size_t k = 1; k <= 40; k += 9) { |
| 7644 | GemmMicrokernelTester() |
| 7645 | .mr(3) |
| 7646 | .nr(16) |
| 7647 | .kr(2) |
| 7648 | .sr(4) |
| 7649 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 7650 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7651 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7652 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7653 | } |
| 7654 | } |
| 7655 | } |
| 7656 | |
| 7657 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_div_16_strided_cn) { |
| 7658 | TEST_REQUIRES_ARM_NEON; |
| 7659 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 7660 | for (size_t k = 1; k <= 40; k += 9) { |
| 7661 | GemmMicrokernelTester() |
| 7662 | .mr(3) |
| 7663 | .nr(16) |
| 7664 | .kr(2) |
| 7665 | .sr(4) |
| 7666 | .m(3) |
| 7667 | .n(n) |
| 7668 | .k(k) |
| 7669 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7670 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7671 | } |
| 7672 | } |
| 7673 | } |
| 7674 | |
| 7675 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_div_16_strided_a) { |
| 7676 | TEST_REQUIRES_ARM_NEON; |
| 7677 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 7678 | for (size_t k = 1; k <= 40; k += 9) { |
| 7679 | GemmMicrokernelTester() |
| 7680 | .mr(3) |
| 7681 | .nr(16) |
| 7682 | .kr(2) |
| 7683 | .sr(4) |
| 7684 | .m(3) |
| 7685 | .n(n) |
| 7686 | .k(k) |
| 7687 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7688 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7689 | } |
| 7690 | } |
| 7691 | } |
| 7692 | |
| 7693 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, n_div_16_subtile) { |
| 7694 | TEST_REQUIRES_ARM_NEON; |
| 7695 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 7696 | for (size_t k = 1; k <= 40; k += 9) { |
| 7697 | for (uint32_t m = 1; m <= 3; m++) { |
| 7698 | GemmMicrokernelTester() |
| 7699 | .mr(3) |
| 7700 | .nr(16) |
| 7701 | .kr(2) |
| 7702 | .sr(4) |
| 7703 | .m(m) |
| 7704 | .n(n) |
| 7705 | .k(k) |
| 7706 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7707 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7708 | } |
| 7709 | } |
| 7710 | } |
| 7711 | } |
| 7712 | |
| 7713 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, strided_cm_subtile) { |
| 7714 | TEST_REQUIRES_ARM_NEON; |
| 7715 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7716 | for (uint32_t n = 1; n <= 16; n++) { |
| 7717 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7718 | GemmMicrokernelTester() |
| 7719 | .mr(3) |
| 7720 | .nr(16) |
| 7721 | .kr(2) |
| 7722 | .sr(4) |
| 7723 | .m(m) |
| 7724 | .n(n) |
| 7725 | .k(k) |
| 7726 | .cm_stride(19) |
| 7727 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7728 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7729 | } |
| 7730 | } |
| 7731 | } |
| 7732 | } |
| 7733 | |
| 7734 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, qmin) { |
| 7735 | TEST_REQUIRES_ARM_NEON; |
| 7736 | GemmMicrokernelTester() |
| 7737 | .mr(3) |
| 7738 | .nr(16) |
| 7739 | .kr(2) |
| 7740 | .sr(4) |
| 7741 | .m(3) |
| 7742 | .n(16) |
| 7743 | .k(8) |
| 7744 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7745 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7746 | } |
| 7747 | |
| 7748 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, qmax) { |
| 7749 | TEST_REQUIRES_ARM_NEON; |
| 7750 | GemmMicrokernelTester() |
| 7751 | .mr(3) |
| 7752 | .nr(16) |
| 7753 | .kr(2) |
| 7754 | .sr(4) |
| 7755 | .m(3) |
| 7756 | .n(16) |
| 7757 | .k(8) |
| 7758 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7759 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7760 | } |
| 7761 | |
| 7762 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2S4__NEON_MULL, strided_cm) { |
| 7763 | TEST_REQUIRES_ARM_NEON; |
| 7764 | GemmMicrokernelTester() |
| 7765 | .mr(3) |
| 7766 | .nr(16) |
| 7767 | .kr(2) |
| 7768 | .sr(4) |
| 7769 | .m(3) |
| 7770 | .n(16) |
| 7771 | .k(8) |
| 7772 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7773 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7774 | } |
| 7775 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 7776 | |
| 7777 | |
| 7778 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 7779 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_eq_8) { |
| 7780 | TEST_REQUIRES_ARM_NEON; |
| 7781 | GemmMicrokernelTester() |
| 7782 | .mr(1) |
| 7783 | .nr(16) |
| 7784 | .kr(4) |
| 7785 | .sr(1) |
| 7786 | .m(1) |
| 7787 | .n(16) |
| 7788 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7789 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7790 | } |
| 7791 | |
| 7792 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, strided_cn) { |
| 7793 | TEST_REQUIRES_ARM_NEON; |
| 7794 | GemmMicrokernelTester() |
| 7795 | .mr(1) |
| 7796 | .nr(16) |
| 7797 | .kr(4) |
| 7798 | .sr(1) |
| 7799 | .m(1) |
| 7800 | .n(16) |
| 7801 | .k(8) |
| 7802 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7803 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7804 | } |
| 7805 | |
| 7806 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_eq_8_strided_a) { |
| 7807 | TEST_REQUIRES_ARM_NEON; |
| 7808 | GemmMicrokernelTester() |
| 7809 | .mr(1) |
| 7810 | .nr(16) |
| 7811 | .kr(4) |
| 7812 | .sr(1) |
| 7813 | .m(1) |
| 7814 | .n(16) |
| 7815 | .k(8) |
| 7816 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7817 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7818 | } |
| 7819 | |
| 7820 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_eq_8_subtile) { |
| 7821 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7822 | for (uint32_t n = 1; n <= 16; n++) { |
| 7823 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7824 | GemmMicrokernelTester() |
| 7825 | .mr(1) |
| 7826 | .nr(16) |
| 7827 | .kr(4) |
| 7828 | .sr(1) |
| 7829 | .m(m) |
| 7830 | .n(n) |
| 7831 | .k(8) |
| 7832 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7833 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7834 | } |
| 7835 | } |
| 7836 | } |
| 7837 | |
| 7838 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_eq_8_subtile_m) { |
| 7839 | TEST_REQUIRES_ARM_NEON; |
| 7840 | for (uint32_t m = 1; m <= 1; m++) { |
| 7841 | GemmMicrokernelTester() |
| 7842 | .mr(1) |
| 7843 | .nr(16) |
| 7844 | .kr(4) |
| 7845 | .sr(1) |
| 7846 | .m(m) |
| 7847 | .n(16) |
| 7848 | .k(8) |
| 7849 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7850 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7851 | } |
| 7852 | } |
| 7853 | |
| 7854 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_eq_8_subtile_n) { |
| 7855 | TEST_REQUIRES_ARM_NEON; |
| 7856 | for (uint32_t n = 1; n <= 16; n++) { |
| 7857 | GemmMicrokernelTester() |
| 7858 | .mr(1) |
| 7859 | .nr(16) |
| 7860 | .kr(4) |
| 7861 | .sr(1) |
| 7862 | .m(1) |
| 7863 | .n(n) |
| 7864 | .k(8) |
| 7865 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7866 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7867 | } |
| 7868 | } |
| 7869 | |
| 7870 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_lt_8) { |
| 7871 | TEST_REQUIRES_ARM_NEON; |
| 7872 | for (size_t k = 1; k < 8; k++) { |
| 7873 | GemmMicrokernelTester() |
| 7874 | .mr(1) |
| 7875 | .nr(16) |
| 7876 | .kr(4) |
| 7877 | .sr(1) |
| 7878 | .m(1) |
| 7879 | .n(16) |
| 7880 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7881 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7882 | } |
| 7883 | } |
| 7884 | |
| 7885 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_lt_8_strided_a) { |
| 7886 | TEST_REQUIRES_ARM_NEON; |
| 7887 | for (size_t k = 1; k < 8; k++) { |
| 7888 | GemmMicrokernelTester() |
| 7889 | .mr(1) |
| 7890 | .nr(16) |
| 7891 | .kr(4) |
| 7892 | .sr(1) |
| 7893 | .m(1) |
| 7894 | .n(16) |
| 7895 | .k(k) |
| 7896 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7897 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7898 | } |
| 7899 | } |
| 7900 | |
| 7901 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_lt_8_subtile) { |
| 7902 | TEST_REQUIRES_ARM_NEON; |
| 7903 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7904 | for (uint32_t n = 1; n <= 16; n++) { |
| 7905 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7906 | GemmMicrokernelTester() |
| 7907 | .mr(1) |
| 7908 | .nr(16) |
| 7909 | .kr(4) |
| 7910 | .sr(1) |
| 7911 | .m(m) |
| 7912 | .n(n) |
| 7913 | .k(k) |
| 7914 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7915 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7916 | } |
| 7917 | } |
| 7918 | } |
| 7919 | } |
| 7920 | |
| 7921 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_gt_8) { |
| 7922 | TEST_REQUIRES_ARM_NEON; |
| 7923 | for (size_t k = 9; k < 16; k++) { |
| 7924 | GemmMicrokernelTester() |
| 7925 | .mr(1) |
| 7926 | .nr(16) |
| 7927 | .kr(4) |
| 7928 | .sr(1) |
| 7929 | .m(1) |
| 7930 | .n(16) |
| 7931 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7932 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7933 | } |
| 7934 | } |
| 7935 | |
| 7936 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_gt_8_strided_a) { |
| 7937 | TEST_REQUIRES_ARM_NEON; |
| 7938 | for (size_t k = 9; k < 16; k++) { |
| 7939 | GemmMicrokernelTester() |
| 7940 | .mr(1) |
| 7941 | .nr(16) |
| 7942 | .kr(4) |
| 7943 | .sr(1) |
| 7944 | .m(1) |
| 7945 | .n(16) |
| 7946 | .k(k) |
| 7947 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7948 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7949 | } |
| 7950 | } |
| 7951 | |
| 7952 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_gt_8_subtile) { |
| 7953 | TEST_REQUIRES_ARM_NEON; |
| 7954 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 7955 | for (uint32_t n = 1; n <= 16; n++) { |
| 7956 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7957 | GemmMicrokernelTester() |
| 7958 | .mr(1) |
| 7959 | .nr(16) |
| 7960 | .kr(4) |
| 7961 | .sr(1) |
| 7962 | .m(m) |
| 7963 | .n(n) |
| 7964 | .k(k) |
| 7965 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7966 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7967 | } |
| 7968 | } |
| 7969 | } |
| 7970 | } |
| 7971 | |
| 7972 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_div_8) { |
| 7973 | TEST_REQUIRES_ARM_NEON; |
| 7974 | for (size_t k = 16; k <= 80; k += 8) { |
| 7975 | GemmMicrokernelTester() |
| 7976 | .mr(1) |
| 7977 | .nr(16) |
| 7978 | .kr(4) |
| 7979 | .sr(1) |
| 7980 | .m(1) |
| 7981 | .n(16) |
| 7982 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7983 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 7984 | } |
| 7985 | } |
| 7986 | |
| 7987 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_div_8_strided_a) { |
| 7988 | TEST_REQUIRES_ARM_NEON; |
| 7989 | for (size_t k = 16; k <= 80; k += 8) { |
| 7990 | GemmMicrokernelTester() |
| 7991 | .mr(1) |
| 7992 | .nr(16) |
| 7993 | .kr(4) |
| 7994 | .sr(1) |
| 7995 | .m(1) |
| 7996 | .n(16) |
| 7997 | .k(k) |
| 7998 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 7999 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8000 | } |
| 8001 | } |
| 8002 | |
| 8003 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, k_div_8_subtile) { |
| 8004 | TEST_REQUIRES_ARM_NEON; |
| 8005 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8006 | for (uint32_t n = 1; n <= 16; n++) { |
| 8007 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8008 | GemmMicrokernelTester() |
| 8009 | .mr(1) |
| 8010 | .nr(16) |
| 8011 | .kr(4) |
| 8012 | .sr(1) |
| 8013 | .m(m) |
| 8014 | .n(n) |
| 8015 | .k(k) |
| 8016 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8017 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8018 | } |
| 8019 | } |
| 8020 | } |
| 8021 | } |
| 8022 | |
| 8023 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_gt_16) { |
| 8024 | TEST_REQUIRES_ARM_NEON; |
| 8025 | for (uint32_t n = 17; n < 32; n++) { |
| 8026 | for (size_t k = 1; k <= 40; k += 9) { |
| 8027 | GemmMicrokernelTester() |
| 8028 | .mr(1) |
| 8029 | .nr(16) |
| 8030 | .kr(4) |
| 8031 | .sr(1) |
| 8032 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8033 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8034 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8035 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8036 | } |
| 8037 | } |
| 8038 | } |
| 8039 | |
| 8040 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_gt_16_strided_cn) { |
| 8041 | TEST_REQUIRES_ARM_NEON; |
| 8042 | for (uint32_t n = 17; n < 32; n++) { |
| 8043 | for (size_t k = 1; k <= 40; k += 9) { |
| 8044 | GemmMicrokernelTester() |
| 8045 | .mr(1) |
| 8046 | .nr(16) |
| 8047 | .kr(4) |
| 8048 | .sr(1) |
| 8049 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8050 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8051 | .k(k) |
| 8052 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8053 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8054 | } |
| 8055 | } |
| 8056 | } |
| 8057 | |
| 8058 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_gt_16_strided_a) { |
| 8059 | TEST_REQUIRES_ARM_NEON; |
| 8060 | for (uint32_t n = 17; n < 32; n++) { |
| 8061 | for (size_t k = 1; k <= 40; k += 9) { |
| 8062 | GemmMicrokernelTester() |
| 8063 | .mr(1) |
| 8064 | .nr(16) |
| 8065 | .kr(4) |
| 8066 | .sr(1) |
| 8067 | .m(1) |
| 8068 | .n(n) |
| 8069 | .k(k) |
| 8070 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8071 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8072 | } |
| 8073 | } |
| 8074 | } |
| 8075 | |
| 8076 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_gt_16_subtile) { |
| 8077 | TEST_REQUIRES_ARM_NEON; |
| 8078 | for (uint32_t n = 17; n < 32; n++) { |
| 8079 | for (size_t k = 1; k <= 40; k += 9) { |
| 8080 | for (uint32_t m = 1; m <= 1; m++) { |
| 8081 | GemmMicrokernelTester() |
| 8082 | .mr(1) |
| 8083 | .nr(16) |
| 8084 | .kr(4) |
| 8085 | .sr(1) |
| 8086 | .m(m) |
| 8087 | .n(n) |
| 8088 | .k(k) |
| 8089 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8090 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8091 | } |
| 8092 | } |
| 8093 | } |
| 8094 | } |
| 8095 | |
| 8096 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_div_16) { |
| 8097 | TEST_REQUIRES_ARM_NEON; |
| 8098 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8099 | for (size_t k = 1; k <= 40; k += 9) { |
| 8100 | GemmMicrokernelTester() |
| 8101 | .mr(1) |
| 8102 | .nr(16) |
| 8103 | .kr(4) |
| 8104 | .sr(1) |
| 8105 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8106 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8107 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8108 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8109 | } |
| 8110 | } |
| 8111 | } |
| 8112 | |
| 8113 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_div_16_strided_cn) { |
| 8114 | TEST_REQUIRES_ARM_NEON; |
| 8115 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8116 | for (size_t k = 1; k <= 40; k += 9) { |
| 8117 | GemmMicrokernelTester() |
| 8118 | .mr(1) |
| 8119 | .nr(16) |
| 8120 | .kr(4) |
| 8121 | .sr(1) |
| 8122 | .m(1) |
| 8123 | .n(n) |
| 8124 | .k(k) |
| 8125 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8126 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8127 | } |
| 8128 | } |
| 8129 | } |
| 8130 | |
| 8131 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_div_16_strided_a) { |
| 8132 | TEST_REQUIRES_ARM_NEON; |
| 8133 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8134 | for (size_t k = 1; k <= 40; k += 9) { |
| 8135 | GemmMicrokernelTester() |
| 8136 | .mr(1) |
| 8137 | .nr(16) |
| 8138 | .kr(4) |
| 8139 | .sr(1) |
| 8140 | .m(1) |
| 8141 | .n(n) |
| 8142 | .k(k) |
| 8143 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8144 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8145 | } |
| 8146 | } |
| 8147 | } |
| 8148 | |
| 8149 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, n_div_16_subtile) { |
| 8150 | TEST_REQUIRES_ARM_NEON; |
| 8151 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 8152 | for (size_t k = 1; k <= 40; k += 9) { |
| 8153 | for (uint32_t m = 1; m <= 1; m++) { |
| 8154 | GemmMicrokernelTester() |
| 8155 | .mr(1) |
| 8156 | .nr(16) |
| 8157 | .kr(4) |
| 8158 | .sr(1) |
| 8159 | .m(m) |
| 8160 | .n(n) |
| 8161 | .k(k) |
| 8162 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8163 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8164 | } |
| 8165 | } |
| 8166 | } |
| 8167 | } |
| 8168 | |
| 8169 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, strided_cm_subtile) { |
| 8170 | TEST_REQUIRES_ARM_NEON; |
| 8171 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8172 | for (uint32_t n = 1; n <= 16; n++) { |
| 8173 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8174 | GemmMicrokernelTester() |
| 8175 | .mr(1) |
| 8176 | .nr(16) |
| 8177 | .kr(4) |
| 8178 | .sr(1) |
| 8179 | .m(m) |
| 8180 | .n(n) |
| 8181 | .k(k) |
| 8182 | .cm_stride(19) |
| 8183 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8184 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8185 | } |
| 8186 | } |
| 8187 | } |
| 8188 | } |
| 8189 | |
| 8190 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, qmin) { |
| 8191 | TEST_REQUIRES_ARM_NEON; |
| 8192 | GemmMicrokernelTester() |
| 8193 | .mr(1) |
| 8194 | .nr(16) |
| 8195 | .kr(4) |
| 8196 | .sr(1) |
| 8197 | .m(1) |
| 8198 | .n(16) |
| 8199 | .k(8) |
| 8200 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8201 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8202 | } |
| 8203 | |
| 8204 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, qmax) { |
| 8205 | TEST_REQUIRES_ARM_NEON; |
| 8206 | GemmMicrokernelTester() |
| 8207 | .mr(1) |
| 8208 | .nr(16) |
| 8209 | .kr(4) |
| 8210 | .sr(1) |
| 8211 | .m(1) |
| 8212 | .n(16) |
| 8213 | .k(8) |
| 8214 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8215 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8216 | } |
| 8217 | |
| 8218 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C4__NEON_MULL_DUP, strided_cm) { |
| 8219 | TEST_REQUIRES_ARM_NEON; |
| 8220 | GemmMicrokernelTester() |
| 8221 | .mr(1) |
| 8222 | .nr(16) |
| 8223 | .kr(4) |
| 8224 | .sr(1) |
| 8225 | .m(1) |
| 8226 | .n(16) |
| 8227 | .k(8) |
| 8228 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8229 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8230 | } |
| 8231 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 8232 | |
| 8233 | |
| 8234 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 8235 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16) { |
| 8236 | TEST_REQUIRES_ARM_NEON; |
| 8237 | GemmMicrokernelTester() |
| 8238 | .mr(3) |
| 8239 | .nr(8) |
| 8240 | .kr(4) |
| 8241 | .sr(1) |
| 8242 | .m(3) |
| 8243 | .n(8) |
| 8244 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8245 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8246 | } |
| 8247 | |
| 8248 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, strided_cn) { |
| 8249 | TEST_REQUIRES_ARM_NEON; |
| 8250 | GemmMicrokernelTester() |
| 8251 | .mr(3) |
| 8252 | .nr(8) |
| 8253 | .kr(4) |
| 8254 | .sr(1) |
| 8255 | .m(3) |
| 8256 | .n(8) |
| 8257 | .k(16) |
| 8258 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8259 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8260 | } |
| 8261 | |
| 8262 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16_strided_a) { |
| 8263 | TEST_REQUIRES_ARM_NEON; |
| 8264 | GemmMicrokernelTester() |
| 8265 | .mr(3) |
| 8266 | .nr(8) |
| 8267 | .kr(4) |
| 8268 | .sr(1) |
| 8269 | .m(3) |
| 8270 | .n(8) |
| 8271 | .k(16) |
| 8272 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8273 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8274 | } |
| 8275 | |
| 8276 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16_subtile) { |
| 8277 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8278 | for (uint32_t n = 1; n <= 8; n++) { |
| 8279 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8280 | GemmMicrokernelTester() |
| 8281 | .mr(3) |
| 8282 | .nr(8) |
| 8283 | .kr(4) |
| 8284 | .sr(1) |
| 8285 | .m(m) |
| 8286 | .n(n) |
| 8287 | .k(16) |
| 8288 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8289 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8290 | } |
| 8291 | } |
| 8292 | } |
| 8293 | |
| 8294 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16_subtile_m) { |
| 8295 | TEST_REQUIRES_ARM_NEON; |
| 8296 | for (uint32_t m = 1; m <= 3; m++) { |
| 8297 | GemmMicrokernelTester() |
| 8298 | .mr(3) |
| 8299 | .nr(8) |
| 8300 | .kr(4) |
| 8301 | .sr(1) |
| 8302 | .m(m) |
| 8303 | .n(8) |
| 8304 | .k(16) |
| 8305 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8306 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8307 | } |
| 8308 | } |
| 8309 | |
| 8310 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16_subtile_n) { |
| 8311 | TEST_REQUIRES_ARM_NEON; |
| 8312 | for (uint32_t n = 1; n <= 8; n++) { |
| 8313 | GemmMicrokernelTester() |
| 8314 | .mr(3) |
| 8315 | .nr(8) |
| 8316 | .kr(4) |
| 8317 | .sr(1) |
| 8318 | .m(3) |
| 8319 | .n(n) |
| 8320 | .k(16) |
| 8321 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8322 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8323 | } |
| 8324 | } |
| 8325 | |
| 8326 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_lt_16) { |
| 8327 | TEST_REQUIRES_ARM_NEON; |
| 8328 | for (size_t k = 1; k < 16; k++) { |
| 8329 | GemmMicrokernelTester() |
| 8330 | .mr(3) |
| 8331 | .nr(8) |
| 8332 | .kr(4) |
| 8333 | .sr(1) |
| 8334 | .m(3) |
| 8335 | .n(8) |
| 8336 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8337 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8338 | } |
| 8339 | } |
| 8340 | |
| 8341 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_lt_16_strided_a) { |
| 8342 | TEST_REQUIRES_ARM_NEON; |
| 8343 | for (size_t k = 1; k < 16; k++) { |
| 8344 | GemmMicrokernelTester() |
| 8345 | .mr(3) |
| 8346 | .nr(8) |
| 8347 | .kr(4) |
| 8348 | .sr(1) |
| 8349 | .m(3) |
| 8350 | .n(8) |
| 8351 | .k(k) |
| 8352 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8353 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8354 | } |
| 8355 | } |
| 8356 | |
| 8357 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_lt_16_subtile) { |
| 8358 | TEST_REQUIRES_ARM_NEON; |
| 8359 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8360 | for (uint32_t n = 1; n <= 8; n++) { |
| 8361 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8362 | GemmMicrokernelTester() |
| 8363 | .mr(3) |
| 8364 | .nr(8) |
| 8365 | .kr(4) |
| 8366 | .sr(1) |
| 8367 | .m(m) |
| 8368 | .n(n) |
| 8369 | .k(k) |
| 8370 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8371 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8372 | } |
| 8373 | } |
| 8374 | } |
| 8375 | } |
| 8376 | |
| 8377 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_gt_16) { |
| 8378 | TEST_REQUIRES_ARM_NEON; |
| 8379 | for (size_t k = 17; k < 32; k++) { |
| 8380 | GemmMicrokernelTester() |
| 8381 | .mr(3) |
| 8382 | .nr(8) |
| 8383 | .kr(4) |
| 8384 | .sr(1) |
| 8385 | .m(3) |
| 8386 | .n(8) |
| 8387 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8388 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8389 | } |
| 8390 | } |
| 8391 | |
| 8392 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_gt_16_strided_a) { |
| 8393 | TEST_REQUIRES_ARM_NEON; |
| 8394 | for (size_t k = 17; k < 32; k++) { |
| 8395 | GemmMicrokernelTester() |
| 8396 | .mr(3) |
| 8397 | .nr(8) |
| 8398 | .kr(4) |
| 8399 | .sr(1) |
| 8400 | .m(3) |
| 8401 | .n(8) |
| 8402 | .k(k) |
| 8403 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8404 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8405 | } |
| 8406 | } |
| 8407 | |
| 8408 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_gt_16_subtile) { |
| 8409 | TEST_REQUIRES_ARM_NEON; |
| 8410 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8411 | for (uint32_t n = 1; n <= 8; n++) { |
| 8412 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8413 | GemmMicrokernelTester() |
| 8414 | .mr(3) |
| 8415 | .nr(8) |
| 8416 | .kr(4) |
| 8417 | .sr(1) |
| 8418 | .m(m) |
| 8419 | .n(n) |
| 8420 | .k(k) |
| 8421 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8422 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8423 | } |
| 8424 | } |
| 8425 | } |
| 8426 | } |
| 8427 | |
| 8428 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_div_16) { |
| 8429 | TEST_REQUIRES_ARM_NEON; |
| 8430 | for (size_t k = 32; k <= 160; k += 16) { |
| 8431 | GemmMicrokernelTester() |
| 8432 | .mr(3) |
| 8433 | .nr(8) |
| 8434 | .kr(4) |
| 8435 | .sr(1) |
| 8436 | .m(3) |
| 8437 | .n(8) |
| 8438 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8439 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8440 | } |
| 8441 | } |
| 8442 | |
| 8443 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_div_16_strided_a) { |
| 8444 | TEST_REQUIRES_ARM_NEON; |
| 8445 | for (size_t k = 32; k <= 160; k += 16) { |
| 8446 | GemmMicrokernelTester() |
| 8447 | .mr(3) |
| 8448 | .nr(8) |
| 8449 | .kr(4) |
| 8450 | .sr(1) |
| 8451 | .m(3) |
| 8452 | .n(8) |
| 8453 | .k(k) |
| 8454 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8455 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8456 | } |
| 8457 | } |
| 8458 | |
| 8459 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_div_16_subtile) { |
| 8460 | TEST_REQUIRES_ARM_NEON; |
| 8461 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8462 | for (uint32_t n = 1; n <= 8; n++) { |
| 8463 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8464 | GemmMicrokernelTester() |
| 8465 | .mr(3) |
| 8466 | .nr(8) |
| 8467 | .kr(4) |
| 8468 | .sr(1) |
| 8469 | .m(m) |
| 8470 | .n(n) |
| 8471 | .k(k) |
| 8472 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8473 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8474 | } |
| 8475 | } |
| 8476 | } |
| 8477 | } |
| 8478 | |
| 8479 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8) { |
| 8480 | TEST_REQUIRES_ARM_NEON; |
| 8481 | for (uint32_t n = 9; n < 16; n++) { |
| 8482 | for (size_t k = 1; k <= 80; k += 17) { |
| 8483 | GemmMicrokernelTester() |
| 8484 | .mr(3) |
| 8485 | .nr(8) |
| 8486 | .kr(4) |
| 8487 | .sr(1) |
| 8488 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8489 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8490 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8491 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8492 | } |
| 8493 | } |
| 8494 | } |
| 8495 | |
| 8496 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8_strided_cn) { |
| 8497 | TEST_REQUIRES_ARM_NEON; |
| 8498 | for (uint32_t n = 9; n < 16; n++) { |
| 8499 | for (size_t k = 1; k <= 80; k += 17) { |
| 8500 | GemmMicrokernelTester() |
| 8501 | .mr(3) |
| 8502 | .nr(8) |
| 8503 | .kr(4) |
| 8504 | .sr(1) |
| 8505 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8506 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8507 | .k(k) |
| 8508 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8509 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8510 | } |
| 8511 | } |
| 8512 | } |
| 8513 | |
| 8514 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8_strided_a) { |
| 8515 | TEST_REQUIRES_ARM_NEON; |
| 8516 | for (uint32_t n = 9; n < 16; n++) { |
| 8517 | for (size_t k = 1; k <= 80; k += 17) { |
| 8518 | GemmMicrokernelTester() |
| 8519 | .mr(3) |
| 8520 | .nr(8) |
| 8521 | .kr(4) |
| 8522 | .sr(1) |
| 8523 | .m(3) |
| 8524 | .n(n) |
| 8525 | .k(k) |
| 8526 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8527 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8528 | } |
| 8529 | } |
| 8530 | } |
| 8531 | |
| 8532 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8_subtile) { |
| 8533 | TEST_REQUIRES_ARM_NEON; |
| 8534 | for (uint32_t n = 9; n < 16; n++) { |
| 8535 | for (size_t k = 1; k <= 80; k += 17) { |
| 8536 | for (uint32_t m = 1; m <= 3; m++) { |
| 8537 | GemmMicrokernelTester() |
| 8538 | .mr(3) |
| 8539 | .nr(8) |
| 8540 | .kr(4) |
| 8541 | .sr(1) |
| 8542 | .m(m) |
| 8543 | .n(n) |
| 8544 | .k(k) |
| 8545 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8546 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8547 | } |
| 8548 | } |
| 8549 | } |
| 8550 | } |
| 8551 | |
| 8552 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8) { |
| 8553 | TEST_REQUIRES_ARM_NEON; |
| 8554 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8555 | for (size_t k = 1; k <= 80; k += 17) { |
| 8556 | GemmMicrokernelTester() |
| 8557 | .mr(3) |
| 8558 | .nr(8) |
| 8559 | .kr(4) |
| 8560 | .sr(1) |
| 8561 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8562 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8563 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8564 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8565 | } |
| 8566 | } |
| 8567 | } |
| 8568 | |
| 8569 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8_strided_cn) { |
| 8570 | TEST_REQUIRES_ARM_NEON; |
| 8571 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8572 | for (size_t k = 1; k <= 80; k += 17) { |
| 8573 | GemmMicrokernelTester() |
| 8574 | .mr(3) |
| 8575 | .nr(8) |
| 8576 | .kr(4) |
| 8577 | .sr(1) |
| 8578 | .m(3) |
| 8579 | .n(n) |
| 8580 | .k(k) |
| 8581 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8582 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8583 | } |
| 8584 | } |
| 8585 | } |
| 8586 | |
| 8587 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8_strided_a) { |
| 8588 | TEST_REQUIRES_ARM_NEON; |
| 8589 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8590 | for (size_t k = 1; k <= 80; k += 17) { |
| 8591 | GemmMicrokernelTester() |
| 8592 | .mr(3) |
| 8593 | .nr(8) |
| 8594 | .kr(4) |
| 8595 | .sr(1) |
| 8596 | .m(3) |
| 8597 | .n(n) |
| 8598 | .k(k) |
| 8599 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8600 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8601 | } |
| 8602 | } |
| 8603 | } |
| 8604 | |
| 8605 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8_subtile) { |
| 8606 | TEST_REQUIRES_ARM_NEON; |
| 8607 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 8608 | for (size_t k = 1; k <= 80; k += 17) { |
| 8609 | for (uint32_t m = 1; m <= 3; m++) { |
| 8610 | GemmMicrokernelTester() |
| 8611 | .mr(3) |
| 8612 | .nr(8) |
| 8613 | .kr(4) |
| 8614 | .sr(1) |
| 8615 | .m(m) |
| 8616 | .n(n) |
| 8617 | .k(k) |
| 8618 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8619 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8620 | } |
| 8621 | } |
| 8622 | } |
| 8623 | } |
| 8624 | |
| 8625 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, strided_cm_subtile) { |
| 8626 | TEST_REQUIRES_ARM_NEON; |
| 8627 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8628 | for (uint32_t n = 1; n <= 8; n++) { |
| 8629 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8630 | GemmMicrokernelTester() |
| 8631 | .mr(3) |
| 8632 | .nr(8) |
| 8633 | .kr(4) |
| 8634 | .sr(1) |
| 8635 | .m(m) |
| 8636 | .n(n) |
| 8637 | .k(k) |
| 8638 | .cm_stride(11) |
| 8639 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8640 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8641 | } |
| 8642 | } |
| 8643 | } |
| 8644 | } |
| 8645 | |
| 8646 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, qmin) { |
| 8647 | TEST_REQUIRES_ARM_NEON; |
| 8648 | GemmMicrokernelTester() |
| 8649 | .mr(3) |
| 8650 | .nr(8) |
| 8651 | .kr(4) |
| 8652 | .sr(1) |
| 8653 | .m(3) |
| 8654 | .n(8) |
| 8655 | .k(16) |
| 8656 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8657 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8658 | } |
| 8659 | |
| 8660 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, qmax) { |
| 8661 | TEST_REQUIRES_ARM_NEON; |
| 8662 | GemmMicrokernelTester() |
| 8663 | .mr(3) |
| 8664 | .nr(8) |
| 8665 | .kr(4) |
| 8666 | .sr(1) |
| 8667 | .m(3) |
| 8668 | .n(8) |
| 8669 | .k(16) |
| 8670 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8671 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8672 | } |
| 8673 | |
| 8674 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, strided_cm) { |
| 8675 | TEST_REQUIRES_ARM_NEON; |
| 8676 | GemmMicrokernelTester() |
| 8677 | .mr(3) |
| 8678 | .nr(8) |
| 8679 | .kr(4) |
| 8680 | .sr(1) |
| 8681 | .m(3) |
| 8682 | .n(8) |
| 8683 | .k(16) |
| 8684 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8685 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8686 | } |
| 8687 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 8688 | |
| 8689 | |
| 8690 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 8691 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_eq_8) { |
| 8692 | TEST_REQUIRES_ARM_NEON; |
| 8693 | GemmMicrokernelTester() |
| 8694 | .mr(2) |
| 8695 | .nr(8) |
| 8696 | .kr(4) |
| 8697 | .sr(1) |
| 8698 | .m(2) |
| 8699 | .n(8) |
| 8700 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8701 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8702 | } |
| 8703 | |
| 8704 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, strided_cn) { |
| 8705 | TEST_REQUIRES_ARM_NEON; |
| 8706 | GemmMicrokernelTester() |
| 8707 | .mr(2) |
| 8708 | .nr(8) |
| 8709 | .kr(4) |
| 8710 | .sr(1) |
| 8711 | .m(2) |
| 8712 | .n(8) |
| 8713 | .k(8) |
| 8714 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8715 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8716 | } |
| 8717 | |
| 8718 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_eq_8_strided_a) { |
| 8719 | TEST_REQUIRES_ARM_NEON; |
| 8720 | GemmMicrokernelTester() |
| 8721 | .mr(2) |
| 8722 | .nr(8) |
| 8723 | .kr(4) |
| 8724 | .sr(1) |
| 8725 | .m(2) |
| 8726 | .n(8) |
| 8727 | .k(8) |
| 8728 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8729 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8730 | } |
| 8731 | |
| 8732 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_eq_8_subtile) { |
| 8733 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8734 | for (uint32_t n = 1; n <= 8; n++) { |
| 8735 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8736 | GemmMicrokernelTester() |
| 8737 | .mr(2) |
| 8738 | .nr(8) |
| 8739 | .kr(4) |
| 8740 | .sr(1) |
| 8741 | .m(m) |
| 8742 | .n(n) |
| 8743 | .k(8) |
| 8744 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8745 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8746 | } |
| 8747 | } |
| 8748 | } |
| 8749 | |
| 8750 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_eq_8_subtile_m) { |
| 8751 | TEST_REQUIRES_ARM_NEON; |
| 8752 | for (uint32_t m = 1; m <= 2; m++) { |
| 8753 | GemmMicrokernelTester() |
| 8754 | .mr(2) |
| 8755 | .nr(8) |
| 8756 | .kr(4) |
| 8757 | .sr(1) |
| 8758 | .m(m) |
| 8759 | .n(8) |
| 8760 | .k(8) |
| 8761 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8762 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8763 | } |
| 8764 | } |
| 8765 | |
| 8766 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_eq_8_subtile_n) { |
| 8767 | TEST_REQUIRES_ARM_NEON; |
| 8768 | for (uint32_t n = 1; n <= 8; n++) { |
| 8769 | GemmMicrokernelTester() |
| 8770 | .mr(2) |
| 8771 | .nr(8) |
| 8772 | .kr(4) |
| 8773 | .sr(1) |
| 8774 | .m(2) |
| 8775 | .n(n) |
| 8776 | .k(8) |
| 8777 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8778 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8779 | } |
| 8780 | } |
| 8781 | |
| 8782 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_lt_8) { |
| 8783 | TEST_REQUIRES_ARM_NEON; |
| 8784 | for (size_t k = 1; k < 8; k++) { |
| 8785 | GemmMicrokernelTester() |
| 8786 | .mr(2) |
| 8787 | .nr(8) |
| 8788 | .kr(4) |
| 8789 | .sr(1) |
| 8790 | .m(2) |
| 8791 | .n(8) |
| 8792 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8793 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8794 | } |
| 8795 | } |
| 8796 | |
| 8797 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_lt_8_strided_a) { |
| 8798 | TEST_REQUIRES_ARM_NEON; |
| 8799 | for (size_t k = 1; k < 8; k++) { |
| 8800 | GemmMicrokernelTester() |
| 8801 | .mr(2) |
| 8802 | .nr(8) |
| 8803 | .kr(4) |
| 8804 | .sr(1) |
| 8805 | .m(2) |
| 8806 | .n(8) |
| 8807 | .k(k) |
| 8808 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8809 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8810 | } |
| 8811 | } |
| 8812 | |
| 8813 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_lt_8_subtile) { |
| 8814 | TEST_REQUIRES_ARM_NEON; |
| 8815 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8816 | for (uint32_t n = 1; n <= 8; n++) { |
| 8817 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8818 | GemmMicrokernelTester() |
| 8819 | .mr(2) |
| 8820 | .nr(8) |
| 8821 | .kr(4) |
| 8822 | .sr(1) |
| 8823 | .m(m) |
| 8824 | .n(n) |
| 8825 | .k(k) |
| 8826 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8827 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8828 | } |
| 8829 | } |
| 8830 | } |
| 8831 | } |
| 8832 | |
| 8833 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_gt_8) { |
| 8834 | TEST_REQUIRES_ARM_NEON; |
| 8835 | for (size_t k = 9; k < 16; k++) { |
| 8836 | GemmMicrokernelTester() |
| 8837 | .mr(2) |
| 8838 | .nr(8) |
| 8839 | .kr(4) |
| 8840 | .sr(1) |
| 8841 | .m(2) |
| 8842 | .n(8) |
| 8843 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8844 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8845 | } |
| 8846 | } |
| 8847 | |
| 8848 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_gt_8_strided_a) { |
| 8849 | TEST_REQUIRES_ARM_NEON; |
| 8850 | for (size_t k = 9; k < 16; k++) { |
| 8851 | GemmMicrokernelTester() |
| 8852 | .mr(2) |
| 8853 | .nr(8) |
| 8854 | .kr(4) |
| 8855 | .sr(1) |
| 8856 | .m(2) |
| 8857 | .n(8) |
| 8858 | .k(k) |
| 8859 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8860 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8861 | } |
| 8862 | } |
| 8863 | |
| 8864 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_gt_8_subtile) { |
| 8865 | TEST_REQUIRES_ARM_NEON; |
| 8866 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8867 | for (uint32_t n = 1; n <= 8; n++) { |
| 8868 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8869 | GemmMicrokernelTester() |
| 8870 | .mr(2) |
| 8871 | .nr(8) |
| 8872 | .kr(4) |
| 8873 | .sr(1) |
| 8874 | .m(m) |
| 8875 | .n(n) |
| 8876 | .k(k) |
| 8877 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8878 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8879 | } |
| 8880 | } |
| 8881 | } |
| 8882 | } |
| 8883 | |
| 8884 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_div_8) { |
| 8885 | TEST_REQUIRES_ARM_NEON; |
| 8886 | for (size_t k = 16; k <= 80; k += 8) { |
| 8887 | GemmMicrokernelTester() |
| 8888 | .mr(2) |
| 8889 | .nr(8) |
| 8890 | .kr(4) |
| 8891 | .sr(1) |
| 8892 | .m(2) |
| 8893 | .n(8) |
| 8894 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8895 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8896 | } |
| 8897 | } |
| 8898 | |
| 8899 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_div_8_strided_a) { |
| 8900 | TEST_REQUIRES_ARM_NEON; |
| 8901 | for (size_t k = 16; k <= 80; k += 8) { |
| 8902 | GemmMicrokernelTester() |
| 8903 | .mr(2) |
| 8904 | .nr(8) |
| 8905 | .kr(4) |
| 8906 | .sr(1) |
| 8907 | .m(2) |
| 8908 | .n(8) |
| 8909 | .k(k) |
| 8910 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8911 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8912 | } |
| 8913 | } |
| 8914 | |
| 8915 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, k_div_8_subtile) { |
| 8916 | TEST_REQUIRES_ARM_NEON; |
| 8917 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 8918 | for (uint32_t n = 1; n <= 8; n++) { |
| 8919 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8920 | GemmMicrokernelTester() |
| 8921 | .mr(2) |
| 8922 | .nr(8) |
| 8923 | .kr(4) |
| 8924 | .sr(1) |
| 8925 | .m(m) |
| 8926 | .n(n) |
| 8927 | .k(k) |
| 8928 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8929 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8930 | } |
| 8931 | } |
| 8932 | } |
| 8933 | } |
| 8934 | |
| 8935 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_gt_8) { |
| 8936 | TEST_REQUIRES_ARM_NEON; |
| 8937 | for (uint32_t n = 9; n < 16; n++) { |
| 8938 | for (size_t k = 1; k <= 40; k += 9) { |
| 8939 | GemmMicrokernelTester() |
| 8940 | .mr(2) |
| 8941 | .nr(8) |
| 8942 | .kr(4) |
| 8943 | .sr(1) |
| 8944 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8945 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8946 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8947 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8948 | } |
| 8949 | } |
| 8950 | } |
| 8951 | |
| 8952 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_gt_8_strided_cn) { |
| 8953 | TEST_REQUIRES_ARM_NEON; |
| 8954 | for (uint32_t n = 9; n < 16; n++) { |
| 8955 | for (size_t k = 1; k <= 40; k += 9) { |
| 8956 | GemmMicrokernelTester() |
| 8957 | .mr(2) |
| 8958 | .nr(8) |
| 8959 | .kr(4) |
| 8960 | .sr(1) |
| 8961 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 8962 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8963 | .k(k) |
| 8964 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8965 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8966 | } |
| 8967 | } |
| 8968 | } |
| 8969 | |
| 8970 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_gt_8_strided_a) { |
| 8971 | TEST_REQUIRES_ARM_NEON; |
| 8972 | for (uint32_t n = 9; n < 16; n++) { |
| 8973 | for (size_t k = 1; k <= 40; k += 9) { |
| 8974 | GemmMicrokernelTester() |
| 8975 | .mr(2) |
| 8976 | .nr(8) |
| 8977 | .kr(4) |
| 8978 | .sr(1) |
| 8979 | .m(2) |
| 8980 | .n(n) |
| 8981 | .k(k) |
| 8982 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 8983 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 8984 | } |
| 8985 | } |
| 8986 | } |
| 8987 | |
| 8988 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_gt_8_subtile) { |
| 8989 | TEST_REQUIRES_ARM_NEON; |
| 8990 | for (uint32_t n = 9; n < 16; n++) { |
| 8991 | for (size_t k = 1; k <= 40; k += 9) { |
| 8992 | for (uint32_t m = 1; m <= 2; m++) { |
| 8993 | GemmMicrokernelTester() |
| 8994 | .mr(2) |
| 8995 | .nr(8) |
| 8996 | .kr(4) |
| 8997 | .sr(1) |
| 8998 | .m(m) |
| 8999 | .n(n) |
| 9000 | .k(k) |
| 9001 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9002 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9003 | } |
| 9004 | } |
| 9005 | } |
| 9006 | } |
| 9007 | |
| 9008 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_div_8) { |
| 9009 | TEST_REQUIRES_ARM_NEON; |
| 9010 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9011 | for (size_t k = 1; k <= 40; k += 9) { |
| 9012 | GemmMicrokernelTester() |
| 9013 | .mr(2) |
| 9014 | .nr(8) |
| 9015 | .kr(4) |
| 9016 | .sr(1) |
| 9017 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9018 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9019 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9020 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9021 | } |
| 9022 | } |
| 9023 | } |
| 9024 | |
| 9025 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_div_8_strided_cn) { |
| 9026 | TEST_REQUIRES_ARM_NEON; |
| 9027 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9028 | for (size_t k = 1; k <= 40; k += 9) { |
| 9029 | GemmMicrokernelTester() |
| 9030 | .mr(2) |
| 9031 | .nr(8) |
| 9032 | .kr(4) |
| 9033 | .sr(1) |
| 9034 | .m(2) |
| 9035 | .n(n) |
| 9036 | .k(k) |
| 9037 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9038 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9039 | } |
| 9040 | } |
| 9041 | } |
| 9042 | |
| 9043 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_div_8_strided_a) { |
| 9044 | TEST_REQUIRES_ARM_NEON; |
| 9045 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9046 | for (size_t k = 1; k <= 40; k += 9) { |
| 9047 | GemmMicrokernelTester() |
| 9048 | .mr(2) |
| 9049 | .nr(8) |
| 9050 | .kr(4) |
| 9051 | .sr(1) |
| 9052 | .m(2) |
| 9053 | .n(n) |
| 9054 | .k(k) |
| 9055 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9056 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9057 | } |
| 9058 | } |
| 9059 | } |
| 9060 | |
| 9061 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, n_div_8_subtile) { |
| 9062 | TEST_REQUIRES_ARM_NEON; |
| 9063 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9064 | for (size_t k = 1; k <= 40; k += 9) { |
| 9065 | for (uint32_t m = 1; m <= 2; m++) { |
| 9066 | GemmMicrokernelTester() |
| 9067 | .mr(2) |
| 9068 | .nr(8) |
| 9069 | .kr(4) |
| 9070 | .sr(1) |
| 9071 | .m(m) |
| 9072 | .n(n) |
| 9073 | .k(k) |
| 9074 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9075 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9076 | } |
| 9077 | } |
| 9078 | } |
| 9079 | } |
| 9080 | |
| 9081 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, strided_cm_subtile) { |
| 9082 | TEST_REQUIRES_ARM_NEON; |
| 9083 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9084 | for (uint32_t n = 1; n <= 8; n++) { |
| 9085 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9086 | GemmMicrokernelTester() |
| 9087 | .mr(2) |
| 9088 | .nr(8) |
| 9089 | .kr(4) |
| 9090 | .sr(1) |
| 9091 | .m(m) |
| 9092 | .n(n) |
| 9093 | .k(k) |
| 9094 | .cm_stride(11) |
| 9095 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9096 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9097 | } |
| 9098 | } |
| 9099 | } |
| 9100 | } |
| 9101 | |
| 9102 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, qmin) { |
| 9103 | TEST_REQUIRES_ARM_NEON; |
| 9104 | GemmMicrokernelTester() |
| 9105 | .mr(2) |
| 9106 | .nr(8) |
| 9107 | .kr(4) |
| 9108 | .sr(1) |
| 9109 | .m(2) |
| 9110 | .n(8) |
| 9111 | .k(8) |
| 9112 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9113 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9114 | } |
| 9115 | |
| 9116 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, qmax) { |
| 9117 | TEST_REQUIRES_ARM_NEON; |
| 9118 | GemmMicrokernelTester() |
| 9119 | .mr(2) |
| 9120 | .nr(8) |
| 9121 | .kr(4) |
| 9122 | .sr(1) |
| 9123 | .m(2) |
| 9124 | .n(8) |
| 9125 | .k(8) |
| 9126 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9127 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9128 | } |
| 9129 | |
| 9130 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD1R, strided_cm) { |
| 9131 | TEST_REQUIRES_ARM_NEON; |
| 9132 | GemmMicrokernelTester() |
| 9133 | .mr(2) |
| 9134 | .nr(8) |
| 9135 | .kr(4) |
| 9136 | .sr(1) |
| 9137 | .m(2) |
| 9138 | .n(8) |
| 9139 | .k(8) |
| 9140 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9141 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9142 | } |
| 9143 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 9144 | |
| 9145 | |
| 9146 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 9147 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_eq_8) { |
| 9148 | TEST_REQUIRES_ARM_NEON; |
| 9149 | GemmMicrokernelTester() |
| 9150 | .mr(4) |
| 9151 | .nr(8) |
| 9152 | .kr(4) |
| 9153 | .sr(1) |
| 9154 | .m(4) |
| 9155 | .n(8) |
| 9156 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9157 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9158 | } |
| 9159 | |
| 9160 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, strided_cn) { |
| 9161 | TEST_REQUIRES_ARM_NEON; |
| 9162 | GemmMicrokernelTester() |
| 9163 | .mr(4) |
| 9164 | .nr(8) |
| 9165 | .kr(4) |
| 9166 | .sr(1) |
| 9167 | .m(4) |
| 9168 | .n(8) |
| 9169 | .k(8) |
| 9170 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9171 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9172 | } |
| 9173 | |
| 9174 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_eq_8_strided_a) { |
| 9175 | TEST_REQUIRES_ARM_NEON; |
| 9176 | GemmMicrokernelTester() |
| 9177 | .mr(4) |
| 9178 | .nr(8) |
| 9179 | .kr(4) |
| 9180 | .sr(1) |
| 9181 | .m(4) |
| 9182 | .n(8) |
| 9183 | .k(8) |
| 9184 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9185 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9186 | } |
| 9187 | |
| 9188 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_eq_8_subtile) { |
| 9189 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9190 | for (uint32_t n = 1; n <= 8; n++) { |
| 9191 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9192 | GemmMicrokernelTester() |
| 9193 | .mr(4) |
| 9194 | .nr(8) |
| 9195 | .kr(4) |
| 9196 | .sr(1) |
| 9197 | .m(m) |
| 9198 | .n(n) |
| 9199 | .k(8) |
| 9200 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9201 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9202 | } |
| 9203 | } |
| 9204 | } |
| 9205 | |
| 9206 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_eq_8_subtile_m) { |
| 9207 | TEST_REQUIRES_ARM_NEON; |
| 9208 | for (uint32_t m = 1; m <= 4; m++) { |
| 9209 | GemmMicrokernelTester() |
| 9210 | .mr(4) |
| 9211 | .nr(8) |
| 9212 | .kr(4) |
| 9213 | .sr(1) |
| 9214 | .m(m) |
| 9215 | .n(8) |
| 9216 | .k(8) |
| 9217 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9218 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9219 | } |
| 9220 | } |
| 9221 | |
| 9222 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_eq_8_subtile_n) { |
| 9223 | TEST_REQUIRES_ARM_NEON; |
| 9224 | for (uint32_t n = 1; n <= 8; n++) { |
| 9225 | GemmMicrokernelTester() |
| 9226 | .mr(4) |
| 9227 | .nr(8) |
| 9228 | .kr(4) |
| 9229 | .sr(1) |
| 9230 | .m(4) |
| 9231 | .n(n) |
| 9232 | .k(8) |
| 9233 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9234 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9235 | } |
| 9236 | } |
| 9237 | |
| 9238 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_lt_8) { |
| 9239 | TEST_REQUIRES_ARM_NEON; |
| 9240 | for (size_t k = 1; k < 8; k++) { |
| 9241 | GemmMicrokernelTester() |
| 9242 | .mr(4) |
| 9243 | .nr(8) |
| 9244 | .kr(4) |
| 9245 | .sr(1) |
| 9246 | .m(4) |
| 9247 | .n(8) |
| 9248 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9249 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9250 | } |
| 9251 | } |
| 9252 | |
| 9253 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_lt_8_strided_a) { |
| 9254 | TEST_REQUIRES_ARM_NEON; |
| 9255 | for (size_t k = 1; k < 8; k++) { |
| 9256 | GemmMicrokernelTester() |
| 9257 | .mr(4) |
| 9258 | .nr(8) |
| 9259 | .kr(4) |
| 9260 | .sr(1) |
| 9261 | .m(4) |
| 9262 | .n(8) |
| 9263 | .k(k) |
| 9264 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9265 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9266 | } |
| 9267 | } |
| 9268 | |
| 9269 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_lt_8_subtile) { |
| 9270 | TEST_REQUIRES_ARM_NEON; |
| 9271 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9272 | for (uint32_t n = 1; n <= 8; n++) { |
| 9273 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9274 | GemmMicrokernelTester() |
| 9275 | .mr(4) |
| 9276 | .nr(8) |
| 9277 | .kr(4) |
| 9278 | .sr(1) |
| 9279 | .m(m) |
| 9280 | .n(n) |
| 9281 | .k(k) |
| 9282 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9283 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9284 | } |
| 9285 | } |
| 9286 | } |
| 9287 | } |
| 9288 | |
| 9289 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_gt_8) { |
| 9290 | TEST_REQUIRES_ARM_NEON; |
| 9291 | for (size_t k = 9; k < 16; k++) { |
| 9292 | GemmMicrokernelTester() |
| 9293 | .mr(4) |
| 9294 | .nr(8) |
| 9295 | .kr(4) |
| 9296 | .sr(1) |
| 9297 | .m(4) |
| 9298 | .n(8) |
| 9299 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9300 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9301 | } |
| 9302 | } |
| 9303 | |
| 9304 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_gt_8_strided_a) { |
| 9305 | TEST_REQUIRES_ARM_NEON; |
| 9306 | for (size_t k = 9; k < 16; k++) { |
| 9307 | GemmMicrokernelTester() |
| 9308 | .mr(4) |
| 9309 | .nr(8) |
| 9310 | .kr(4) |
| 9311 | .sr(1) |
| 9312 | .m(4) |
| 9313 | .n(8) |
| 9314 | .k(k) |
| 9315 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9316 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9317 | } |
| 9318 | } |
| 9319 | |
| 9320 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_gt_8_subtile) { |
| 9321 | TEST_REQUIRES_ARM_NEON; |
| 9322 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9323 | for (uint32_t n = 1; n <= 8; n++) { |
| 9324 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9325 | GemmMicrokernelTester() |
| 9326 | .mr(4) |
| 9327 | .nr(8) |
| 9328 | .kr(4) |
| 9329 | .sr(1) |
| 9330 | .m(m) |
| 9331 | .n(n) |
| 9332 | .k(k) |
| 9333 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9334 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9335 | } |
| 9336 | } |
| 9337 | } |
| 9338 | } |
| 9339 | |
| 9340 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_div_8) { |
| 9341 | TEST_REQUIRES_ARM_NEON; |
| 9342 | for (size_t k = 16; k <= 80; k += 8) { |
| 9343 | GemmMicrokernelTester() |
| 9344 | .mr(4) |
| 9345 | .nr(8) |
| 9346 | .kr(4) |
| 9347 | .sr(1) |
| 9348 | .m(4) |
| 9349 | .n(8) |
| 9350 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9351 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9352 | } |
| 9353 | } |
| 9354 | |
| 9355 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_div_8_strided_a) { |
| 9356 | TEST_REQUIRES_ARM_NEON; |
| 9357 | for (size_t k = 16; k <= 80; k += 8) { |
| 9358 | GemmMicrokernelTester() |
| 9359 | .mr(4) |
| 9360 | .nr(8) |
| 9361 | .kr(4) |
| 9362 | .sr(1) |
| 9363 | .m(4) |
| 9364 | .n(8) |
| 9365 | .k(k) |
| 9366 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9367 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9368 | } |
| 9369 | } |
| 9370 | |
| 9371 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, k_div_8_subtile) { |
| 9372 | TEST_REQUIRES_ARM_NEON; |
| 9373 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9374 | for (uint32_t n = 1; n <= 8; n++) { |
| 9375 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9376 | GemmMicrokernelTester() |
| 9377 | .mr(4) |
| 9378 | .nr(8) |
| 9379 | .kr(4) |
| 9380 | .sr(1) |
| 9381 | .m(m) |
| 9382 | .n(n) |
| 9383 | .k(k) |
| 9384 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9385 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9386 | } |
| 9387 | } |
| 9388 | } |
| 9389 | } |
| 9390 | |
| 9391 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_gt_8) { |
| 9392 | TEST_REQUIRES_ARM_NEON; |
| 9393 | for (uint32_t n = 9; n < 16; n++) { |
| 9394 | for (size_t k = 1; k <= 40; k += 9) { |
| 9395 | GemmMicrokernelTester() |
| 9396 | .mr(4) |
| 9397 | .nr(8) |
| 9398 | .kr(4) |
| 9399 | .sr(1) |
| 9400 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9401 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9402 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9403 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9404 | } |
| 9405 | } |
| 9406 | } |
| 9407 | |
| 9408 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_gt_8_strided_cn) { |
| 9409 | TEST_REQUIRES_ARM_NEON; |
| 9410 | for (uint32_t n = 9; n < 16; n++) { |
| 9411 | for (size_t k = 1; k <= 40; k += 9) { |
| 9412 | GemmMicrokernelTester() |
| 9413 | .mr(4) |
| 9414 | .nr(8) |
| 9415 | .kr(4) |
| 9416 | .sr(1) |
| 9417 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9418 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9419 | .k(k) |
| 9420 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9421 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9422 | } |
| 9423 | } |
| 9424 | } |
| 9425 | |
| 9426 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_gt_8_strided_a) { |
| 9427 | TEST_REQUIRES_ARM_NEON; |
| 9428 | for (uint32_t n = 9; n < 16; n++) { |
| 9429 | for (size_t k = 1; k <= 40; k += 9) { |
| 9430 | GemmMicrokernelTester() |
| 9431 | .mr(4) |
| 9432 | .nr(8) |
| 9433 | .kr(4) |
| 9434 | .sr(1) |
| 9435 | .m(4) |
| 9436 | .n(n) |
| 9437 | .k(k) |
| 9438 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9439 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9440 | } |
| 9441 | } |
| 9442 | } |
| 9443 | |
| 9444 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_gt_8_subtile) { |
| 9445 | TEST_REQUIRES_ARM_NEON; |
| 9446 | for (uint32_t n = 9; n < 16; n++) { |
| 9447 | for (size_t k = 1; k <= 40; k += 9) { |
| 9448 | for (uint32_t m = 1; m <= 4; m++) { |
| 9449 | GemmMicrokernelTester() |
| 9450 | .mr(4) |
| 9451 | .nr(8) |
| 9452 | .kr(4) |
| 9453 | .sr(1) |
| 9454 | .m(m) |
| 9455 | .n(n) |
| 9456 | .k(k) |
| 9457 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9458 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9459 | } |
| 9460 | } |
| 9461 | } |
| 9462 | } |
| 9463 | |
| 9464 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_div_8) { |
| 9465 | TEST_REQUIRES_ARM_NEON; |
| 9466 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9467 | for (size_t k = 1; k <= 40; k += 9) { |
| 9468 | GemmMicrokernelTester() |
| 9469 | .mr(4) |
| 9470 | .nr(8) |
| 9471 | .kr(4) |
| 9472 | .sr(1) |
| 9473 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9474 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9475 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9476 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9477 | } |
| 9478 | } |
| 9479 | } |
| 9480 | |
| 9481 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_div_8_strided_cn) { |
| 9482 | TEST_REQUIRES_ARM_NEON; |
| 9483 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9484 | for (size_t k = 1; k <= 40; k += 9) { |
| 9485 | GemmMicrokernelTester() |
| 9486 | .mr(4) |
| 9487 | .nr(8) |
| 9488 | .kr(4) |
| 9489 | .sr(1) |
| 9490 | .m(4) |
| 9491 | .n(n) |
| 9492 | .k(k) |
| 9493 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9494 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9495 | } |
| 9496 | } |
| 9497 | } |
| 9498 | |
| 9499 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_div_8_strided_a) { |
| 9500 | TEST_REQUIRES_ARM_NEON; |
| 9501 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9502 | for (size_t k = 1; k <= 40; k += 9) { |
| 9503 | GemmMicrokernelTester() |
| 9504 | .mr(4) |
| 9505 | .nr(8) |
| 9506 | .kr(4) |
| 9507 | .sr(1) |
| 9508 | .m(4) |
| 9509 | .n(n) |
| 9510 | .k(k) |
| 9511 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9512 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9513 | } |
| 9514 | } |
| 9515 | } |
| 9516 | |
| 9517 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, n_div_8_subtile) { |
| 9518 | TEST_REQUIRES_ARM_NEON; |
| 9519 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 9520 | for (size_t k = 1; k <= 40; k += 9) { |
| 9521 | for (uint32_t m = 1; m <= 4; m++) { |
| 9522 | GemmMicrokernelTester() |
| 9523 | .mr(4) |
| 9524 | .nr(8) |
| 9525 | .kr(4) |
| 9526 | .sr(1) |
| 9527 | .m(m) |
| 9528 | .n(n) |
| 9529 | .k(k) |
| 9530 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9531 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9532 | } |
| 9533 | } |
| 9534 | } |
| 9535 | } |
| 9536 | |
| 9537 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, strided_cm_subtile) { |
| 9538 | TEST_REQUIRES_ARM_NEON; |
| 9539 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9540 | for (uint32_t n = 1; n <= 8; n++) { |
| 9541 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9542 | GemmMicrokernelTester() |
| 9543 | .mr(4) |
| 9544 | .nr(8) |
| 9545 | .kr(4) |
| 9546 | .sr(1) |
| 9547 | .m(m) |
| 9548 | .n(n) |
| 9549 | .k(k) |
| 9550 | .cm_stride(11) |
| 9551 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9552 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9553 | } |
| 9554 | } |
| 9555 | } |
| 9556 | } |
| 9557 | |
| 9558 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, qmin) { |
| 9559 | TEST_REQUIRES_ARM_NEON; |
| 9560 | GemmMicrokernelTester() |
| 9561 | .mr(4) |
| 9562 | .nr(8) |
| 9563 | .kr(4) |
| 9564 | .sr(1) |
| 9565 | .m(4) |
| 9566 | .n(8) |
| 9567 | .k(8) |
| 9568 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9569 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9570 | } |
| 9571 | |
| 9572 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, qmax) { |
| 9573 | TEST_REQUIRES_ARM_NEON; |
| 9574 | GemmMicrokernelTester() |
| 9575 | .mr(4) |
| 9576 | .nr(8) |
| 9577 | .kr(4) |
| 9578 | .sr(1) |
| 9579 | .m(4) |
| 9580 | .n(8) |
| 9581 | .k(8) |
| 9582 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9583 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9584 | } |
| 9585 | |
| 9586 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_LD1R, strided_cm) { |
| 9587 | TEST_REQUIRES_ARM_NEON; |
| 9588 | GemmMicrokernelTester() |
| 9589 | .mr(4) |
| 9590 | .nr(8) |
| 9591 | .kr(4) |
| 9592 | .sr(1) |
| 9593 | .m(4) |
| 9594 | .n(8) |
| 9595 | .k(8) |
| 9596 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9597 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9598 | } |
| 9599 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 9600 | |
| 9601 | |
| 9602 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 9603 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_eq_8) { |
| 9604 | TEST_REQUIRES_ARM_NEON; |
| 9605 | GemmMicrokernelTester() |
| 9606 | .mr(3) |
| 9607 | .nr(16) |
| 9608 | .kr(4) |
| 9609 | .sr(1) |
| 9610 | .m(3) |
| 9611 | .n(16) |
| 9612 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9613 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9614 | } |
| 9615 | |
| 9616 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, strided_cn) { |
| 9617 | TEST_REQUIRES_ARM_NEON; |
| 9618 | GemmMicrokernelTester() |
| 9619 | .mr(3) |
| 9620 | .nr(16) |
| 9621 | .kr(4) |
| 9622 | .sr(1) |
| 9623 | .m(3) |
| 9624 | .n(16) |
| 9625 | .k(8) |
| 9626 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9627 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9628 | } |
| 9629 | |
| 9630 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_eq_8_strided_a) { |
| 9631 | TEST_REQUIRES_ARM_NEON; |
| 9632 | GemmMicrokernelTester() |
| 9633 | .mr(3) |
| 9634 | .nr(16) |
| 9635 | .kr(4) |
| 9636 | .sr(1) |
| 9637 | .m(3) |
| 9638 | .n(16) |
| 9639 | .k(8) |
| 9640 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9641 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9642 | } |
| 9643 | |
| 9644 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_eq_8_subtile) { |
| 9645 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9646 | for (uint32_t n = 1; n <= 16; n++) { |
| 9647 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9648 | GemmMicrokernelTester() |
| 9649 | .mr(3) |
| 9650 | .nr(16) |
| 9651 | .kr(4) |
| 9652 | .sr(1) |
| 9653 | .m(m) |
| 9654 | .n(n) |
| 9655 | .k(8) |
| 9656 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9657 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9658 | } |
| 9659 | } |
| 9660 | } |
| 9661 | |
| 9662 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_eq_8_subtile_m) { |
| 9663 | TEST_REQUIRES_ARM_NEON; |
| 9664 | for (uint32_t m = 1; m <= 3; m++) { |
| 9665 | GemmMicrokernelTester() |
| 9666 | .mr(3) |
| 9667 | .nr(16) |
| 9668 | .kr(4) |
| 9669 | .sr(1) |
| 9670 | .m(m) |
| 9671 | .n(16) |
| 9672 | .k(8) |
| 9673 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9674 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9675 | } |
| 9676 | } |
| 9677 | |
| 9678 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_eq_8_subtile_n) { |
| 9679 | TEST_REQUIRES_ARM_NEON; |
| 9680 | for (uint32_t n = 1; n <= 16; n++) { |
| 9681 | GemmMicrokernelTester() |
| 9682 | .mr(3) |
| 9683 | .nr(16) |
| 9684 | .kr(4) |
| 9685 | .sr(1) |
| 9686 | .m(3) |
| 9687 | .n(n) |
| 9688 | .k(8) |
| 9689 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9690 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9691 | } |
| 9692 | } |
| 9693 | |
| 9694 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_lt_8) { |
| 9695 | TEST_REQUIRES_ARM_NEON; |
| 9696 | for (size_t k = 1; k < 8; k++) { |
| 9697 | GemmMicrokernelTester() |
| 9698 | .mr(3) |
| 9699 | .nr(16) |
| 9700 | .kr(4) |
| 9701 | .sr(1) |
| 9702 | .m(3) |
| 9703 | .n(16) |
| 9704 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9705 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9706 | } |
| 9707 | } |
| 9708 | |
| 9709 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_lt_8_strided_a) { |
| 9710 | TEST_REQUIRES_ARM_NEON; |
| 9711 | for (size_t k = 1; k < 8; k++) { |
| 9712 | GemmMicrokernelTester() |
| 9713 | .mr(3) |
| 9714 | .nr(16) |
| 9715 | .kr(4) |
| 9716 | .sr(1) |
| 9717 | .m(3) |
| 9718 | .n(16) |
| 9719 | .k(k) |
| 9720 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9721 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9722 | } |
| 9723 | } |
| 9724 | |
| 9725 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_lt_8_subtile) { |
| 9726 | TEST_REQUIRES_ARM_NEON; |
| 9727 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9728 | for (uint32_t n = 1; n <= 16; n++) { |
| 9729 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9730 | GemmMicrokernelTester() |
| 9731 | .mr(3) |
| 9732 | .nr(16) |
| 9733 | .kr(4) |
| 9734 | .sr(1) |
| 9735 | .m(m) |
| 9736 | .n(n) |
| 9737 | .k(k) |
| 9738 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9739 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9740 | } |
| 9741 | } |
| 9742 | } |
| 9743 | } |
| 9744 | |
| 9745 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_gt_8) { |
| 9746 | TEST_REQUIRES_ARM_NEON; |
| 9747 | for (size_t k = 9; k < 16; k++) { |
| 9748 | GemmMicrokernelTester() |
| 9749 | .mr(3) |
| 9750 | .nr(16) |
| 9751 | .kr(4) |
| 9752 | .sr(1) |
| 9753 | .m(3) |
| 9754 | .n(16) |
| 9755 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9756 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9757 | } |
| 9758 | } |
| 9759 | |
| 9760 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_gt_8_strided_a) { |
| 9761 | TEST_REQUIRES_ARM_NEON; |
| 9762 | for (size_t k = 9; k < 16; k++) { |
| 9763 | GemmMicrokernelTester() |
| 9764 | .mr(3) |
| 9765 | .nr(16) |
| 9766 | .kr(4) |
| 9767 | .sr(1) |
| 9768 | .m(3) |
| 9769 | .n(16) |
| 9770 | .k(k) |
| 9771 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9772 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9773 | } |
| 9774 | } |
| 9775 | |
| 9776 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_gt_8_subtile) { |
| 9777 | TEST_REQUIRES_ARM_NEON; |
| 9778 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9779 | for (uint32_t n = 1; n <= 16; n++) { |
| 9780 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9781 | GemmMicrokernelTester() |
| 9782 | .mr(3) |
| 9783 | .nr(16) |
| 9784 | .kr(4) |
| 9785 | .sr(1) |
| 9786 | .m(m) |
| 9787 | .n(n) |
| 9788 | .k(k) |
| 9789 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9790 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9791 | } |
| 9792 | } |
| 9793 | } |
| 9794 | } |
| 9795 | |
| 9796 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_div_8) { |
| 9797 | TEST_REQUIRES_ARM_NEON; |
| 9798 | for (size_t k = 16; k <= 80; k += 8) { |
| 9799 | GemmMicrokernelTester() |
| 9800 | .mr(3) |
| 9801 | .nr(16) |
| 9802 | .kr(4) |
| 9803 | .sr(1) |
| 9804 | .m(3) |
| 9805 | .n(16) |
| 9806 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9807 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9808 | } |
| 9809 | } |
| 9810 | |
| 9811 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_div_8_strided_a) { |
| 9812 | TEST_REQUIRES_ARM_NEON; |
| 9813 | for (size_t k = 16; k <= 80; k += 8) { |
| 9814 | GemmMicrokernelTester() |
| 9815 | .mr(3) |
| 9816 | .nr(16) |
| 9817 | .kr(4) |
| 9818 | .sr(1) |
| 9819 | .m(3) |
| 9820 | .n(16) |
| 9821 | .k(k) |
| 9822 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9823 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9824 | } |
| 9825 | } |
| 9826 | |
| 9827 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, k_div_8_subtile) { |
| 9828 | TEST_REQUIRES_ARM_NEON; |
| 9829 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9830 | for (uint32_t n = 1; n <= 16; n++) { |
| 9831 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9832 | GemmMicrokernelTester() |
| 9833 | .mr(3) |
| 9834 | .nr(16) |
| 9835 | .kr(4) |
| 9836 | .sr(1) |
| 9837 | .m(m) |
| 9838 | .n(n) |
| 9839 | .k(k) |
| 9840 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9841 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9842 | } |
| 9843 | } |
| 9844 | } |
| 9845 | } |
| 9846 | |
| 9847 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_gt_16) { |
| 9848 | TEST_REQUIRES_ARM_NEON; |
| 9849 | for (uint32_t n = 17; n < 32; n++) { |
| 9850 | for (size_t k = 1; k <= 40; k += 9) { |
| 9851 | GemmMicrokernelTester() |
| 9852 | .mr(3) |
| 9853 | .nr(16) |
| 9854 | .kr(4) |
| 9855 | .sr(1) |
| 9856 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9857 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9858 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9859 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9860 | } |
| 9861 | } |
| 9862 | } |
| 9863 | |
| 9864 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_gt_16_strided_cn) { |
| 9865 | TEST_REQUIRES_ARM_NEON; |
| 9866 | for (uint32_t n = 17; n < 32; n++) { |
| 9867 | for (size_t k = 1; k <= 40; k += 9) { |
| 9868 | GemmMicrokernelTester() |
| 9869 | .mr(3) |
| 9870 | .nr(16) |
| 9871 | .kr(4) |
| 9872 | .sr(1) |
| 9873 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9874 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9875 | .k(k) |
| 9876 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9877 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9878 | } |
| 9879 | } |
| 9880 | } |
| 9881 | |
| 9882 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_gt_16_strided_a) { |
| 9883 | TEST_REQUIRES_ARM_NEON; |
| 9884 | for (uint32_t n = 17; n < 32; n++) { |
| 9885 | for (size_t k = 1; k <= 40; k += 9) { |
| 9886 | GemmMicrokernelTester() |
| 9887 | .mr(3) |
| 9888 | .nr(16) |
| 9889 | .kr(4) |
| 9890 | .sr(1) |
| 9891 | .m(3) |
| 9892 | .n(n) |
| 9893 | .k(k) |
| 9894 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9895 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9896 | } |
| 9897 | } |
| 9898 | } |
| 9899 | |
| 9900 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_gt_16_subtile) { |
| 9901 | TEST_REQUIRES_ARM_NEON; |
| 9902 | for (uint32_t n = 17; n < 32; n++) { |
| 9903 | for (size_t k = 1; k <= 40; k += 9) { |
| 9904 | for (uint32_t m = 1; m <= 3; m++) { |
| 9905 | GemmMicrokernelTester() |
| 9906 | .mr(3) |
| 9907 | .nr(16) |
| 9908 | .kr(4) |
| 9909 | .sr(1) |
| 9910 | .m(m) |
| 9911 | .n(n) |
| 9912 | .k(k) |
| 9913 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9914 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9915 | } |
| 9916 | } |
| 9917 | } |
| 9918 | } |
| 9919 | |
| 9920 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_div_16) { |
| 9921 | TEST_REQUIRES_ARM_NEON; |
| 9922 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9923 | for (size_t k = 1; k <= 40; k += 9) { |
| 9924 | GemmMicrokernelTester() |
| 9925 | .mr(3) |
| 9926 | .nr(16) |
| 9927 | .kr(4) |
| 9928 | .sr(1) |
| 9929 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 9930 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9931 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9932 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9933 | } |
| 9934 | } |
| 9935 | } |
| 9936 | |
| 9937 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_div_16_strided_cn) { |
| 9938 | TEST_REQUIRES_ARM_NEON; |
| 9939 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9940 | for (size_t k = 1; k <= 40; k += 9) { |
| 9941 | GemmMicrokernelTester() |
| 9942 | .mr(3) |
| 9943 | .nr(16) |
| 9944 | .kr(4) |
| 9945 | .sr(1) |
| 9946 | .m(3) |
| 9947 | .n(n) |
| 9948 | .k(k) |
| 9949 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9950 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9951 | } |
| 9952 | } |
| 9953 | } |
| 9954 | |
| 9955 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_div_16_strided_a) { |
| 9956 | TEST_REQUIRES_ARM_NEON; |
| 9957 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9958 | for (size_t k = 1; k <= 40; k += 9) { |
| 9959 | GemmMicrokernelTester() |
| 9960 | .mr(3) |
| 9961 | .nr(16) |
| 9962 | .kr(4) |
| 9963 | .sr(1) |
| 9964 | .m(3) |
| 9965 | .n(n) |
| 9966 | .k(k) |
| 9967 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9968 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9969 | } |
| 9970 | } |
| 9971 | } |
| 9972 | |
| 9973 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, n_div_16_subtile) { |
| 9974 | TEST_REQUIRES_ARM_NEON; |
| 9975 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 9976 | for (size_t k = 1; k <= 40; k += 9) { |
| 9977 | for (uint32_t m = 1; m <= 3; m++) { |
| 9978 | GemmMicrokernelTester() |
| 9979 | .mr(3) |
| 9980 | .nr(16) |
| 9981 | .kr(4) |
| 9982 | .sr(1) |
| 9983 | .m(m) |
| 9984 | .n(n) |
| 9985 | .k(k) |
| 9986 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 9987 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9988 | } |
| 9989 | } |
| 9990 | } |
| 9991 | } |
| 9992 | |
| 9993 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, strided_cm_subtile) { |
| 9994 | TEST_REQUIRES_ARM_NEON; |
| 9995 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 9996 | for (uint32_t n = 1; n <= 16; n++) { |
| 9997 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 9998 | GemmMicrokernelTester() |
| 9999 | .mr(3) |
| 10000 | .nr(16) |
| 10001 | .kr(4) |
| 10002 | .sr(1) |
| 10003 | .m(m) |
| 10004 | .n(n) |
| 10005 | .k(k) |
| 10006 | .cm_stride(19) |
| 10007 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10008 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10009 | } |
| 10010 | } |
| 10011 | } |
| 10012 | } |
| 10013 | |
| 10014 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, qmin) { |
| 10015 | TEST_REQUIRES_ARM_NEON; |
| 10016 | GemmMicrokernelTester() |
| 10017 | .mr(3) |
| 10018 | .nr(16) |
| 10019 | .kr(4) |
| 10020 | .sr(1) |
| 10021 | .m(3) |
| 10022 | .n(16) |
| 10023 | .k(8) |
| 10024 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10025 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10026 | } |
| 10027 | |
| 10028 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, qmax) { |
| 10029 | TEST_REQUIRES_ARM_NEON; |
| 10030 | GemmMicrokernelTester() |
| 10031 | .mr(3) |
| 10032 | .nr(16) |
| 10033 | .kr(4) |
| 10034 | .sr(1) |
| 10035 | .m(3) |
| 10036 | .n(16) |
| 10037 | .k(8) |
| 10038 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10039 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10040 | } |
| 10041 | |
| 10042 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C4__NEON_MULL_LD1R, strided_cm) { |
| 10043 | TEST_REQUIRES_ARM_NEON; |
| 10044 | GemmMicrokernelTester() |
| 10045 | .mr(3) |
| 10046 | .nr(16) |
| 10047 | .kr(4) |
| 10048 | .sr(1) |
| 10049 | .m(3) |
| 10050 | .n(16) |
| 10051 | .k(8) |
| 10052 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10053 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10054 | } |
| 10055 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 10056 | |
| 10057 | |
| 10058 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 10059 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_eq_8) { |
| 10060 | TEST_REQUIRES_ARM_NEON; |
| 10061 | GemmMicrokernelTester() |
| 10062 | .mr(3) |
| 10063 | .nr(8) |
| 10064 | .kr(4) |
| 10065 | .sr(1) |
| 10066 | .m(3) |
| 10067 | .n(8) |
| 10068 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10069 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10070 | } |
| 10071 | |
| 10072 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, strided_cn) { |
| 10073 | TEST_REQUIRES_ARM_NEON; |
| 10074 | GemmMicrokernelTester() |
| 10075 | .mr(3) |
| 10076 | .nr(8) |
| 10077 | .kr(4) |
| 10078 | .sr(1) |
| 10079 | .m(3) |
| 10080 | .n(8) |
| 10081 | .k(8) |
| 10082 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10083 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10084 | } |
| 10085 | |
| 10086 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_eq_8_strided_a) { |
| 10087 | TEST_REQUIRES_ARM_NEON; |
| 10088 | GemmMicrokernelTester() |
| 10089 | .mr(3) |
| 10090 | .nr(8) |
| 10091 | .kr(4) |
| 10092 | .sr(1) |
| 10093 | .m(3) |
| 10094 | .n(8) |
| 10095 | .k(8) |
| 10096 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10097 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10098 | } |
| 10099 | |
| 10100 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_eq_8_subtile) { |
| 10101 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10102 | for (uint32_t n = 1; n <= 8; n++) { |
| 10103 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10104 | GemmMicrokernelTester() |
| 10105 | .mr(3) |
| 10106 | .nr(8) |
| 10107 | .kr(4) |
| 10108 | .sr(1) |
| 10109 | .m(m) |
| 10110 | .n(n) |
| 10111 | .k(8) |
| 10112 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10113 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10114 | } |
| 10115 | } |
| 10116 | } |
| 10117 | |
| 10118 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_eq_8_subtile_m) { |
| 10119 | TEST_REQUIRES_ARM_NEON; |
| 10120 | for (uint32_t m = 1; m <= 3; m++) { |
| 10121 | GemmMicrokernelTester() |
| 10122 | .mr(3) |
| 10123 | .nr(8) |
| 10124 | .kr(4) |
| 10125 | .sr(1) |
| 10126 | .m(m) |
| 10127 | .n(8) |
| 10128 | .k(8) |
| 10129 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10130 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10131 | } |
| 10132 | } |
| 10133 | |
| 10134 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_eq_8_subtile_n) { |
| 10135 | TEST_REQUIRES_ARM_NEON; |
| 10136 | for (uint32_t n = 1; n <= 8; n++) { |
| 10137 | GemmMicrokernelTester() |
| 10138 | .mr(3) |
| 10139 | .nr(8) |
| 10140 | .kr(4) |
| 10141 | .sr(1) |
| 10142 | .m(3) |
| 10143 | .n(n) |
| 10144 | .k(8) |
| 10145 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10146 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10147 | } |
| 10148 | } |
| 10149 | |
| 10150 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_lt_8) { |
| 10151 | TEST_REQUIRES_ARM_NEON; |
| 10152 | for (size_t k = 1; k < 8; k++) { |
| 10153 | GemmMicrokernelTester() |
| 10154 | .mr(3) |
| 10155 | .nr(8) |
| 10156 | .kr(4) |
| 10157 | .sr(1) |
| 10158 | .m(3) |
| 10159 | .n(8) |
| 10160 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10161 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10162 | } |
| 10163 | } |
| 10164 | |
| 10165 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_lt_8_strided_a) { |
| 10166 | TEST_REQUIRES_ARM_NEON; |
| 10167 | for (size_t k = 1; k < 8; k++) { |
| 10168 | GemmMicrokernelTester() |
| 10169 | .mr(3) |
| 10170 | .nr(8) |
| 10171 | .kr(4) |
| 10172 | .sr(1) |
| 10173 | .m(3) |
| 10174 | .n(8) |
| 10175 | .k(k) |
| 10176 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10177 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10178 | } |
| 10179 | } |
| 10180 | |
| 10181 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_lt_8_subtile) { |
| 10182 | TEST_REQUIRES_ARM_NEON; |
| 10183 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10184 | for (uint32_t n = 1; n <= 8; n++) { |
| 10185 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10186 | GemmMicrokernelTester() |
| 10187 | .mr(3) |
| 10188 | .nr(8) |
| 10189 | .kr(4) |
| 10190 | .sr(1) |
| 10191 | .m(m) |
| 10192 | .n(n) |
| 10193 | .k(k) |
| 10194 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10195 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10196 | } |
| 10197 | } |
| 10198 | } |
| 10199 | } |
| 10200 | |
| 10201 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_gt_8) { |
| 10202 | TEST_REQUIRES_ARM_NEON; |
| 10203 | for (size_t k = 9; k < 16; k++) { |
| 10204 | GemmMicrokernelTester() |
| 10205 | .mr(3) |
| 10206 | .nr(8) |
| 10207 | .kr(4) |
| 10208 | .sr(1) |
| 10209 | .m(3) |
| 10210 | .n(8) |
| 10211 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10212 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10213 | } |
| 10214 | } |
| 10215 | |
| 10216 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_gt_8_strided_a) { |
| 10217 | TEST_REQUIRES_ARM_NEON; |
| 10218 | for (size_t k = 9; k < 16; k++) { |
| 10219 | GemmMicrokernelTester() |
| 10220 | .mr(3) |
| 10221 | .nr(8) |
| 10222 | .kr(4) |
| 10223 | .sr(1) |
| 10224 | .m(3) |
| 10225 | .n(8) |
| 10226 | .k(k) |
| 10227 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10228 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10229 | } |
| 10230 | } |
| 10231 | |
| 10232 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_gt_8_subtile) { |
| 10233 | TEST_REQUIRES_ARM_NEON; |
| 10234 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10235 | for (uint32_t n = 1; n <= 8; n++) { |
| 10236 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10237 | GemmMicrokernelTester() |
| 10238 | .mr(3) |
| 10239 | .nr(8) |
| 10240 | .kr(4) |
| 10241 | .sr(1) |
| 10242 | .m(m) |
| 10243 | .n(n) |
| 10244 | .k(k) |
| 10245 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10246 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10247 | } |
| 10248 | } |
| 10249 | } |
| 10250 | } |
| 10251 | |
| 10252 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_div_8) { |
| 10253 | TEST_REQUIRES_ARM_NEON; |
| 10254 | for (size_t k = 16; k <= 80; k += 8) { |
| 10255 | GemmMicrokernelTester() |
| 10256 | .mr(3) |
| 10257 | .nr(8) |
| 10258 | .kr(4) |
| 10259 | .sr(1) |
| 10260 | .m(3) |
| 10261 | .n(8) |
| 10262 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10263 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10264 | } |
| 10265 | } |
| 10266 | |
| 10267 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_div_8_strided_a) { |
| 10268 | TEST_REQUIRES_ARM_NEON; |
| 10269 | for (size_t k = 16; k <= 80; k += 8) { |
| 10270 | GemmMicrokernelTester() |
| 10271 | .mr(3) |
| 10272 | .nr(8) |
| 10273 | .kr(4) |
| 10274 | .sr(1) |
| 10275 | .m(3) |
| 10276 | .n(8) |
| 10277 | .k(k) |
| 10278 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10279 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10280 | } |
| 10281 | } |
| 10282 | |
| 10283 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, k_div_8_subtile) { |
| 10284 | TEST_REQUIRES_ARM_NEON; |
| 10285 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10286 | for (uint32_t n = 1; n <= 8; n++) { |
| 10287 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10288 | GemmMicrokernelTester() |
| 10289 | .mr(3) |
| 10290 | .nr(8) |
| 10291 | .kr(4) |
| 10292 | .sr(1) |
| 10293 | .m(m) |
| 10294 | .n(n) |
| 10295 | .k(k) |
| 10296 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10297 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10298 | } |
| 10299 | } |
| 10300 | } |
| 10301 | } |
| 10302 | |
| 10303 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_gt_8) { |
| 10304 | TEST_REQUIRES_ARM_NEON; |
| 10305 | for (uint32_t n = 9; n < 16; n++) { |
| 10306 | for (size_t k = 1; k <= 40; k += 9) { |
| 10307 | GemmMicrokernelTester() |
| 10308 | .mr(3) |
| 10309 | .nr(8) |
| 10310 | .kr(4) |
| 10311 | .sr(1) |
| 10312 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10313 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10314 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10315 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10316 | } |
| 10317 | } |
| 10318 | } |
| 10319 | |
| 10320 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_gt_8_strided_cn) { |
| 10321 | TEST_REQUIRES_ARM_NEON; |
| 10322 | for (uint32_t n = 9; n < 16; n++) { |
| 10323 | for (size_t k = 1; k <= 40; k += 9) { |
| 10324 | GemmMicrokernelTester() |
| 10325 | .mr(3) |
| 10326 | .nr(8) |
| 10327 | .kr(4) |
| 10328 | .sr(1) |
| 10329 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10330 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10331 | .k(k) |
| 10332 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10333 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10334 | } |
| 10335 | } |
| 10336 | } |
| 10337 | |
| 10338 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_gt_8_strided_a) { |
| 10339 | TEST_REQUIRES_ARM_NEON; |
| 10340 | for (uint32_t n = 9; n < 16; n++) { |
| 10341 | for (size_t k = 1; k <= 40; k += 9) { |
| 10342 | GemmMicrokernelTester() |
| 10343 | .mr(3) |
| 10344 | .nr(8) |
| 10345 | .kr(4) |
| 10346 | .sr(1) |
| 10347 | .m(3) |
| 10348 | .n(n) |
| 10349 | .k(k) |
| 10350 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10351 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10352 | } |
| 10353 | } |
| 10354 | } |
| 10355 | |
| 10356 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_gt_8_subtile) { |
| 10357 | TEST_REQUIRES_ARM_NEON; |
| 10358 | for (uint32_t n = 9; n < 16; n++) { |
| 10359 | for (size_t k = 1; k <= 40; k += 9) { |
| 10360 | for (uint32_t m = 1; m <= 3; m++) { |
| 10361 | GemmMicrokernelTester() |
| 10362 | .mr(3) |
| 10363 | .nr(8) |
| 10364 | .kr(4) |
| 10365 | .sr(1) |
| 10366 | .m(m) |
| 10367 | .n(n) |
| 10368 | .k(k) |
| 10369 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10370 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10371 | } |
| 10372 | } |
| 10373 | } |
| 10374 | } |
| 10375 | |
| 10376 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_div_8) { |
| 10377 | TEST_REQUIRES_ARM_NEON; |
| 10378 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 10379 | for (size_t k = 1; k <= 40; k += 9) { |
| 10380 | GemmMicrokernelTester() |
| 10381 | .mr(3) |
| 10382 | .nr(8) |
| 10383 | .kr(4) |
| 10384 | .sr(1) |
| 10385 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10386 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10387 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10388 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10389 | } |
| 10390 | } |
| 10391 | } |
| 10392 | |
| 10393 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_div_8_strided_cn) { |
| 10394 | TEST_REQUIRES_ARM_NEON; |
| 10395 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 10396 | for (size_t k = 1; k <= 40; k += 9) { |
| 10397 | GemmMicrokernelTester() |
| 10398 | .mr(3) |
| 10399 | .nr(8) |
| 10400 | .kr(4) |
| 10401 | .sr(1) |
| 10402 | .m(3) |
| 10403 | .n(n) |
| 10404 | .k(k) |
| 10405 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10406 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10407 | } |
| 10408 | } |
| 10409 | } |
| 10410 | |
| 10411 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_div_8_strided_a) { |
| 10412 | TEST_REQUIRES_ARM_NEON; |
| 10413 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 10414 | for (size_t k = 1; k <= 40; k += 9) { |
| 10415 | GemmMicrokernelTester() |
| 10416 | .mr(3) |
| 10417 | .nr(8) |
| 10418 | .kr(4) |
| 10419 | .sr(1) |
| 10420 | .m(3) |
| 10421 | .n(n) |
| 10422 | .k(k) |
| 10423 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10424 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10425 | } |
| 10426 | } |
| 10427 | } |
| 10428 | |
| 10429 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, n_div_8_subtile) { |
| 10430 | TEST_REQUIRES_ARM_NEON; |
| 10431 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 10432 | for (size_t k = 1; k <= 40; k += 9) { |
| 10433 | for (uint32_t m = 1; m <= 3; m++) { |
| 10434 | GemmMicrokernelTester() |
| 10435 | .mr(3) |
| 10436 | .nr(8) |
| 10437 | .kr(4) |
| 10438 | .sr(1) |
| 10439 | .m(m) |
| 10440 | .n(n) |
| 10441 | .k(k) |
| 10442 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10443 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10444 | } |
| 10445 | } |
| 10446 | } |
| 10447 | } |
| 10448 | |
| 10449 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, strided_cm_subtile) { |
| 10450 | TEST_REQUIRES_ARM_NEON; |
| 10451 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10452 | for (uint32_t n = 1; n <= 8; n++) { |
| 10453 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10454 | GemmMicrokernelTester() |
| 10455 | .mr(3) |
| 10456 | .nr(8) |
| 10457 | .kr(4) |
| 10458 | .sr(1) |
| 10459 | .m(m) |
| 10460 | .n(n) |
| 10461 | .k(k) |
| 10462 | .cm_stride(11) |
| 10463 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10464 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10465 | } |
| 10466 | } |
| 10467 | } |
| 10468 | } |
| 10469 | |
| 10470 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, qmin) { |
| 10471 | TEST_REQUIRES_ARM_NEON; |
| 10472 | GemmMicrokernelTester() |
| 10473 | .mr(3) |
| 10474 | .nr(8) |
| 10475 | .kr(4) |
| 10476 | .sr(1) |
| 10477 | .m(3) |
| 10478 | .n(8) |
| 10479 | .k(8) |
| 10480 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10481 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10482 | } |
| 10483 | |
| 10484 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, qmax) { |
| 10485 | TEST_REQUIRES_ARM_NEON; |
| 10486 | GemmMicrokernelTester() |
| 10487 | .mr(3) |
| 10488 | .nr(8) |
| 10489 | .kr(4) |
| 10490 | .sr(1) |
| 10491 | .m(3) |
| 10492 | .n(8) |
| 10493 | .k(8) |
| 10494 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10495 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10496 | } |
| 10497 | |
| 10498 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C4__NEON_MULL_LD2R, strided_cm) { |
| 10499 | TEST_REQUIRES_ARM_NEON; |
| 10500 | GemmMicrokernelTester() |
| 10501 | .mr(3) |
| 10502 | .nr(8) |
| 10503 | .kr(4) |
| 10504 | .sr(1) |
| 10505 | .m(3) |
| 10506 | .n(8) |
| 10507 | .k(8) |
| 10508 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10509 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10510 | } |
| 10511 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 10512 | |
| 10513 | |
| 10514 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 10515 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_eq_16) { |
| 10516 | TEST_REQUIRES_ARM_NEON; |
| 10517 | GemmMicrokernelTester() |
| 10518 | .mr(2) |
| 10519 | .nr(16) |
| 10520 | .kr(4) |
| 10521 | .sr(1) |
| 10522 | .m(2) |
| 10523 | .n(16) |
| 10524 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10525 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10526 | } |
| 10527 | |
| 10528 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, strided_cn) { |
| 10529 | TEST_REQUIRES_ARM_NEON; |
| 10530 | GemmMicrokernelTester() |
| 10531 | .mr(2) |
| 10532 | .nr(16) |
| 10533 | .kr(4) |
| 10534 | .sr(1) |
| 10535 | .m(2) |
| 10536 | .n(16) |
| 10537 | .k(16) |
| 10538 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10539 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10540 | } |
| 10541 | |
| 10542 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_eq_16_strided_a) { |
| 10543 | TEST_REQUIRES_ARM_NEON; |
| 10544 | GemmMicrokernelTester() |
| 10545 | .mr(2) |
| 10546 | .nr(16) |
| 10547 | .kr(4) |
| 10548 | .sr(1) |
| 10549 | .m(2) |
| 10550 | .n(16) |
| 10551 | .k(16) |
| 10552 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10553 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10554 | } |
| 10555 | |
| 10556 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_eq_16_subtile) { |
| 10557 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10558 | for (uint32_t n = 1; n <= 16; n++) { |
| 10559 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10560 | GemmMicrokernelTester() |
| 10561 | .mr(2) |
| 10562 | .nr(16) |
| 10563 | .kr(4) |
| 10564 | .sr(1) |
| 10565 | .m(m) |
| 10566 | .n(n) |
| 10567 | .k(16) |
| 10568 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10569 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10570 | } |
| 10571 | } |
| 10572 | } |
| 10573 | |
| 10574 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_eq_16_subtile_m) { |
| 10575 | TEST_REQUIRES_ARM_NEON; |
| 10576 | for (uint32_t m = 1; m <= 2; m++) { |
| 10577 | GemmMicrokernelTester() |
| 10578 | .mr(2) |
| 10579 | .nr(16) |
| 10580 | .kr(4) |
| 10581 | .sr(1) |
| 10582 | .m(m) |
| 10583 | .n(16) |
| 10584 | .k(16) |
| 10585 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10586 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10587 | } |
| 10588 | } |
| 10589 | |
| 10590 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_eq_16_subtile_n) { |
| 10591 | TEST_REQUIRES_ARM_NEON; |
| 10592 | for (uint32_t n = 1; n <= 16; n++) { |
| 10593 | GemmMicrokernelTester() |
| 10594 | .mr(2) |
| 10595 | .nr(16) |
| 10596 | .kr(4) |
| 10597 | .sr(1) |
| 10598 | .m(2) |
| 10599 | .n(n) |
| 10600 | .k(16) |
| 10601 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10602 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10603 | } |
| 10604 | } |
| 10605 | |
| 10606 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_lt_16) { |
| 10607 | TEST_REQUIRES_ARM_NEON; |
| 10608 | for (size_t k = 1; k < 16; k++) { |
| 10609 | GemmMicrokernelTester() |
| 10610 | .mr(2) |
| 10611 | .nr(16) |
| 10612 | .kr(4) |
| 10613 | .sr(1) |
| 10614 | .m(2) |
| 10615 | .n(16) |
| 10616 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10617 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10618 | } |
| 10619 | } |
| 10620 | |
| 10621 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_lt_16_strided_a) { |
| 10622 | TEST_REQUIRES_ARM_NEON; |
| 10623 | for (size_t k = 1; k < 16; k++) { |
| 10624 | GemmMicrokernelTester() |
| 10625 | .mr(2) |
| 10626 | .nr(16) |
| 10627 | .kr(4) |
| 10628 | .sr(1) |
| 10629 | .m(2) |
| 10630 | .n(16) |
| 10631 | .k(k) |
| 10632 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10633 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10634 | } |
| 10635 | } |
| 10636 | |
| 10637 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_lt_16_subtile) { |
| 10638 | TEST_REQUIRES_ARM_NEON; |
| 10639 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10640 | for (uint32_t n = 1; n <= 16; n++) { |
| 10641 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10642 | GemmMicrokernelTester() |
| 10643 | .mr(2) |
| 10644 | .nr(16) |
| 10645 | .kr(4) |
| 10646 | .sr(1) |
| 10647 | .m(m) |
| 10648 | .n(n) |
| 10649 | .k(k) |
| 10650 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10651 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10652 | } |
| 10653 | } |
| 10654 | } |
| 10655 | } |
| 10656 | |
| 10657 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_gt_16) { |
| 10658 | TEST_REQUIRES_ARM_NEON; |
| 10659 | for (size_t k = 17; k < 32; k++) { |
| 10660 | GemmMicrokernelTester() |
| 10661 | .mr(2) |
| 10662 | .nr(16) |
| 10663 | .kr(4) |
| 10664 | .sr(1) |
| 10665 | .m(2) |
| 10666 | .n(16) |
| 10667 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10668 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10669 | } |
| 10670 | } |
| 10671 | |
| 10672 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_gt_16_strided_a) { |
| 10673 | TEST_REQUIRES_ARM_NEON; |
| 10674 | for (size_t k = 17; k < 32; k++) { |
| 10675 | GemmMicrokernelTester() |
| 10676 | .mr(2) |
| 10677 | .nr(16) |
| 10678 | .kr(4) |
| 10679 | .sr(1) |
| 10680 | .m(2) |
| 10681 | .n(16) |
| 10682 | .k(k) |
| 10683 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10684 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10685 | } |
| 10686 | } |
| 10687 | |
| 10688 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_gt_16_subtile) { |
| 10689 | TEST_REQUIRES_ARM_NEON; |
| 10690 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10691 | for (uint32_t n = 1; n <= 16; n++) { |
| 10692 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10693 | GemmMicrokernelTester() |
| 10694 | .mr(2) |
| 10695 | .nr(16) |
| 10696 | .kr(4) |
| 10697 | .sr(1) |
| 10698 | .m(m) |
| 10699 | .n(n) |
| 10700 | .k(k) |
| 10701 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10702 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10703 | } |
| 10704 | } |
| 10705 | } |
| 10706 | } |
| 10707 | |
| 10708 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_div_16) { |
| 10709 | TEST_REQUIRES_ARM_NEON; |
| 10710 | for (size_t k = 32; k <= 160; k += 16) { |
| 10711 | GemmMicrokernelTester() |
| 10712 | .mr(2) |
| 10713 | .nr(16) |
| 10714 | .kr(4) |
| 10715 | .sr(1) |
| 10716 | .m(2) |
| 10717 | .n(16) |
| 10718 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10719 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10720 | } |
| 10721 | } |
| 10722 | |
| 10723 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_div_16_strided_a) { |
| 10724 | TEST_REQUIRES_ARM_NEON; |
| 10725 | for (size_t k = 32; k <= 160; k += 16) { |
| 10726 | GemmMicrokernelTester() |
| 10727 | .mr(2) |
| 10728 | .nr(16) |
| 10729 | .kr(4) |
| 10730 | .sr(1) |
| 10731 | .m(2) |
| 10732 | .n(16) |
| 10733 | .k(k) |
| 10734 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10735 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10736 | } |
| 10737 | } |
| 10738 | |
| 10739 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, k_div_16_subtile) { |
| 10740 | TEST_REQUIRES_ARM_NEON; |
| 10741 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10742 | for (uint32_t n = 1; n <= 16; n++) { |
| 10743 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10744 | GemmMicrokernelTester() |
| 10745 | .mr(2) |
| 10746 | .nr(16) |
| 10747 | .kr(4) |
| 10748 | .sr(1) |
| 10749 | .m(m) |
| 10750 | .n(n) |
| 10751 | .k(k) |
| 10752 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10753 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10754 | } |
| 10755 | } |
| 10756 | } |
| 10757 | } |
| 10758 | |
| 10759 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_gt_16) { |
| 10760 | TEST_REQUIRES_ARM_NEON; |
| 10761 | for (uint32_t n = 17; n < 32; n++) { |
| 10762 | for (size_t k = 1; k <= 80; k += 17) { |
| 10763 | GemmMicrokernelTester() |
| 10764 | .mr(2) |
| 10765 | .nr(16) |
| 10766 | .kr(4) |
| 10767 | .sr(1) |
| 10768 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10769 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10770 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10771 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10772 | } |
| 10773 | } |
| 10774 | } |
| 10775 | |
| 10776 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_gt_16_strided_cn) { |
| 10777 | TEST_REQUIRES_ARM_NEON; |
| 10778 | for (uint32_t n = 17; n < 32; n++) { |
| 10779 | for (size_t k = 1; k <= 80; k += 17) { |
| 10780 | GemmMicrokernelTester() |
| 10781 | .mr(2) |
| 10782 | .nr(16) |
| 10783 | .kr(4) |
| 10784 | .sr(1) |
| 10785 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10786 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10787 | .k(k) |
| 10788 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10789 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10790 | } |
| 10791 | } |
| 10792 | } |
| 10793 | |
| 10794 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_gt_16_strided_a) { |
| 10795 | TEST_REQUIRES_ARM_NEON; |
| 10796 | for (uint32_t n = 17; n < 32; n++) { |
| 10797 | for (size_t k = 1; k <= 80; k += 17) { |
| 10798 | GemmMicrokernelTester() |
| 10799 | .mr(2) |
| 10800 | .nr(16) |
| 10801 | .kr(4) |
| 10802 | .sr(1) |
| 10803 | .m(2) |
| 10804 | .n(n) |
| 10805 | .k(k) |
| 10806 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10807 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10808 | } |
| 10809 | } |
| 10810 | } |
| 10811 | |
| 10812 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_gt_16_subtile) { |
| 10813 | TEST_REQUIRES_ARM_NEON; |
| 10814 | for (uint32_t n = 17; n < 32; n++) { |
| 10815 | for (size_t k = 1; k <= 80; k += 17) { |
| 10816 | for (uint32_t m = 1; m <= 2; m++) { |
| 10817 | GemmMicrokernelTester() |
| 10818 | .mr(2) |
| 10819 | .nr(16) |
| 10820 | .kr(4) |
| 10821 | .sr(1) |
| 10822 | .m(m) |
| 10823 | .n(n) |
| 10824 | .k(k) |
| 10825 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10826 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10827 | } |
| 10828 | } |
| 10829 | } |
| 10830 | } |
| 10831 | |
| 10832 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_div_16) { |
| 10833 | TEST_REQUIRES_ARM_NEON; |
| 10834 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10835 | for (size_t k = 1; k <= 80; k += 17) { |
| 10836 | GemmMicrokernelTester() |
| 10837 | .mr(2) |
| 10838 | .nr(16) |
| 10839 | .kr(4) |
| 10840 | .sr(1) |
| 10841 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 10842 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10843 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10844 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10845 | } |
| 10846 | } |
| 10847 | } |
| 10848 | |
| 10849 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_div_16_strided_cn) { |
| 10850 | TEST_REQUIRES_ARM_NEON; |
| 10851 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10852 | for (size_t k = 1; k <= 80; k += 17) { |
| 10853 | GemmMicrokernelTester() |
| 10854 | .mr(2) |
| 10855 | .nr(16) |
| 10856 | .kr(4) |
| 10857 | .sr(1) |
| 10858 | .m(2) |
| 10859 | .n(n) |
| 10860 | .k(k) |
| 10861 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10862 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10863 | } |
| 10864 | } |
| 10865 | } |
| 10866 | |
| 10867 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_div_16_strided_a) { |
| 10868 | TEST_REQUIRES_ARM_NEON; |
| 10869 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10870 | for (size_t k = 1; k <= 80; k += 17) { |
| 10871 | GemmMicrokernelTester() |
| 10872 | .mr(2) |
| 10873 | .nr(16) |
| 10874 | .kr(4) |
| 10875 | .sr(1) |
| 10876 | .m(2) |
| 10877 | .n(n) |
| 10878 | .k(k) |
| 10879 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10880 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10881 | } |
| 10882 | } |
| 10883 | } |
| 10884 | |
| 10885 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, n_div_16_subtile) { |
| 10886 | TEST_REQUIRES_ARM_NEON; |
| 10887 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 10888 | for (size_t k = 1; k <= 80; k += 17) { |
| 10889 | for (uint32_t m = 1; m <= 2; m++) { |
| 10890 | GemmMicrokernelTester() |
| 10891 | .mr(2) |
| 10892 | .nr(16) |
| 10893 | .kr(4) |
| 10894 | .sr(1) |
| 10895 | .m(m) |
| 10896 | .n(n) |
| 10897 | .k(k) |
| 10898 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10899 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10900 | } |
| 10901 | } |
| 10902 | } |
| 10903 | } |
| 10904 | |
| 10905 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, strided_cm_subtile) { |
| 10906 | TEST_REQUIRES_ARM_NEON; |
| 10907 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 10908 | for (uint32_t n = 1; n <= 16; n++) { |
| 10909 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10910 | GemmMicrokernelTester() |
| 10911 | .mr(2) |
| 10912 | .nr(16) |
| 10913 | .kr(4) |
| 10914 | .sr(1) |
| 10915 | .m(m) |
| 10916 | .n(n) |
| 10917 | .k(k) |
| 10918 | .cm_stride(19) |
| 10919 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10920 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10921 | } |
| 10922 | } |
| 10923 | } |
| 10924 | } |
| 10925 | |
| 10926 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, qmin) { |
| 10927 | TEST_REQUIRES_ARM_NEON; |
| 10928 | GemmMicrokernelTester() |
| 10929 | .mr(2) |
| 10930 | .nr(16) |
| 10931 | .kr(4) |
| 10932 | .sr(1) |
| 10933 | .m(2) |
| 10934 | .n(16) |
| 10935 | .k(16) |
| 10936 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10937 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10938 | } |
| 10939 | |
| 10940 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, qmax) { |
| 10941 | TEST_REQUIRES_ARM_NEON; |
| 10942 | GemmMicrokernelTester() |
| 10943 | .mr(2) |
| 10944 | .nr(16) |
| 10945 | .kr(4) |
| 10946 | .sr(1) |
| 10947 | .m(2) |
| 10948 | .n(16) |
| 10949 | .k(16) |
| 10950 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10951 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10952 | } |
| 10953 | |
| 10954 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16C4__NEON_MLAL_LD2R, strided_cm) { |
| 10955 | TEST_REQUIRES_ARM_NEON; |
| 10956 | GemmMicrokernelTester() |
| 10957 | .mr(2) |
| 10958 | .nr(16) |
| 10959 | .kr(4) |
| 10960 | .sr(1) |
| 10961 | .m(2) |
| 10962 | .n(16) |
| 10963 | .k(16) |
| 10964 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10965 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10966 | } |
| 10967 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 10968 | |
| 10969 | |
| 10970 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 10971 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8) { |
| 10972 | TEST_REQUIRES_ARM_NEON; |
| 10973 | GemmMicrokernelTester() |
| 10974 | .mr(4) |
| 10975 | .nr(8) |
| 10976 | .kr(2) |
| 10977 | .sr(1) |
| 10978 | .m(4) |
| 10979 | .n(8) |
| 10980 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10981 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10982 | } |
| 10983 | |
| 10984 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cn) { |
| 10985 | TEST_REQUIRES_ARM_NEON; |
| 10986 | GemmMicrokernelTester() |
| 10987 | .mr(4) |
| 10988 | .nr(8) |
| 10989 | .kr(2) |
| 10990 | .sr(1) |
| 10991 | .m(4) |
| 10992 | .n(8) |
| 10993 | .k(8) |
| 10994 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 10995 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 10996 | } |
| 10997 | |
| 10998 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_strided_a) { |
| 10999 | TEST_REQUIRES_ARM_NEON; |
| 11000 | GemmMicrokernelTester() |
| 11001 | .mr(4) |
| 11002 | .nr(8) |
| 11003 | .kr(2) |
| 11004 | .sr(1) |
| 11005 | .m(4) |
| 11006 | .n(8) |
| 11007 | .k(8) |
| 11008 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11009 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11010 | } |
| 11011 | |
| 11012 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile) { |
| 11013 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11014 | for (uint32_t n = 1; n <= 8; n++) { |
| 11015 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11016 | GemmMicrokernelTester() |
| 11017 | .mr(4) |
| 11018 | .nr(8) |
| 11019 | .kr(2) |
| 11020 | .sr(1) |
| 11021 | .m(m) |
| 11022 | .n(n) |
| 11023 | .k(8) |
| 11024 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11025 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11026 | } |
| 11027 | } |
| 11028 | } |
| 11029 | |
| 11030 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile_m) { |
| 11031 | TEST_REQUIRES_ARM_NEON; |
| 11032 | for (uint32_t m = 1; m <= 4; m++) { |
| 11033 | GemmMicrokernelTester() |
| 11034 | .mr(4) |
| 11035 | .nr(8) |
| 11036 | .kr(2) |
| 11037 | .sr(1) |
| 11038 | .m(m) |
| 11039 | .n(8) |
| 11040 | .k(8) |
| 11041 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11042 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11043 | } |
| 11044 | } |
| 11045 | |
| 11046 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile_n) { |
| 11047 | TEST_REQUIRES_ARM_NEON; |
| 11048 | for (uint32_t n = 1; n <= 8; n++) { |
| 11049 | GemmMicrokernelTester() |
| 11050 | .mr(4) |
| 11051 | .nr(8) |
| 11052 | .kr(2) |
| 11053 | .sr(1) |
| 11054 | .m(4) |
| 11055 | .n(n) |
| 11056 | .k(8) |
| 11057 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11058 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11059 | } |
| 11060 | } |
| 11061 | |
| 11062 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_lt_8) { |
| 11063 | TEST_REQUIRES_ARM_NEON; |
| 11064 | for (size_t k = 1; k < 8; k++) { |
| 11065 | GemmMicrokernelTester() |
| 11066 | .mr(4) |
| 11067 | .nr(8) |
| 11068 | .kr(2) |
| 11069 | .sr(1) |
| 11070 | .m(4) |
| 11071 | .n(8) |
| 11072 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11073 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11074 | } |
| 11075 | } |
| 11076 | |
| 11077 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_lt_8_strided_a) { |
| 11078 | TEST_REQUIRES_ARM_NEON; |
| 11079 | for (size_t k = 1; k < 8; k++) { |
| 11080 | GemmMicrokernelTester() |
| 11081 | .mr(4) |
| 11082 | .nr(8) |
| 11083 | .kr(2) |
| 11084 | .sr(1) |
| 11085 | .m(4) |
| 11086 | .n(8) |
| 11087 | .k(k) |
| 11088 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11089 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11090 | } |
| 11091 | } |
| 11092 | |
| 11093 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_lt_8_subtile) { |
| 11094 | TEST_REQUIRES_ARM_NEON; |
| 11095 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11096 | for (uint32_t n = 1; n <= 8; n++) { |
| 11097 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11098 | GemmMicrokernelTester() |
| 11099 | .mr(4) |
| 11100 | .nr(8) |
| 11101 | .kr(2) |
| 11102 | .sr(1) |
| 11103 | .m(m) |
| 11104 | .n(n) |
| 11105 | .k(k) |
| 11106 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11107 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11108 | } |
| 11109 | } |
| 11110 | } |
| 11111 | } |
| 11112 | |
| 11113 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_gt_8) { |
| 11114 | TEST_REQUIRES_ARM_NEON; |
| 11115 | for (size_t k = 9; k < 16; k++) { |
| 11116 | GemmMicrokernelTester() |
| 11117 | .mr(4) |
| 11118 | .nr(8) |
| 11119 | .kr(2) |
| 11120 | .sr(1) |
| 11121 | .m(4) |
| 11122 | .n(8) |
| 11123 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11124 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11125 | } |
| 11126 | } |
| 11127 | |
| 11128 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_gt_8_strided_a) { |
| 11129 | TEST_REQUIRES_ARM_NEON; |
| 11130 | for (size_t k = 9; k < 16; k++) { |
| 11131 | GemmMicrokernelTester() |
| 11132 | .mr(4) |
| 11133 | .nr(8) |
| 11134 | .kr(2) |
| 11135 | .sr(1) |
| 11136 | .m(4) |
| 11137 | .n(8) |
| 11138 | .k(k) |
| 11139 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11140 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11141 | } |
| 11142 | } |
| 11143 | |
| 11144 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_gt_8_subtile) { |
| 11145 | TEST_REQUIRES_ARM_NEON; |
| 11146 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11147 | for (uint32_t n = 1; n <= 8; n++) { |
| 11148 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11149 | GemmMicrokernelTester() |
| 11150 | .mr(4) |
| 11151 | .nr(8) |
| 11152 | .kr(2) |
| 11153 | .sr(1) |
| 11154 | .m(m) |
| 11155 | .n(n) |
| 11156 | .k(k) |
| 11157 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11158 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11159 | } |
| 11160 | } |
| 11161 | } |
| 11162 | } |
| 11163 | |
| 11164 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_div_8) { |
| 11165 | TEST_REQUIRES_ARM_NEON; |
| 11166 | for (size_t k = 16; k <= 80; k += 8) { |
| 11167 | GemmMicrokernelTester() |
| 11168 | .mr(4) |
| 11169 | .nr(8) |
| 11170 | .kr(2) |
| 11171 | .sr(1) |
| 11172 | .m(4) |
| 11173 | .n(8) |
| 11174 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11175 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11176 | } |
| 11177 | } |
| 11178 | |
| 11179 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_div_8_strided_a) { |
| 11180 | TEST_REQUIRES_ARM_NEON; |
| 11181 | for (size_t k = 16; k <= 80; k += 8) { |
| 11182 | GemmMicrokernelTester() |
| 11183 | .mr(4) |
| 11184 | .nr(8) |
| 11185 | .kr(2) |
| 11186 | .sr(1) |
| 11187 | .m(4) |
| 11188 | .n(8) |
| 11189 | .k(k) |
| 11190 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11191 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11192 | } |
| 11193 | } |
| 11194 | |
| 11195 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_div_8_subtile) { |
| 11196 | TEST_REQUIRES_ARM_NEON; |
| 11197 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11198 | for (uint32_t n = 1; n <= 8; n++) { |
| 11199 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11200 | GemmMicrokernelTester() |
| 11201 | .mr(4) |
| 11202 | .nr(8) |
| 11203 | .kr(2) |
| 11204 | .sr(1) |
| 11205 | .m(m) |
| 11206 | .n(n) |
| 11207 | .k(k) |
| 11208 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11209 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11210 | } |
| 11211 | } |
| 11212 | } |
| 11213 | } |
| 11214 | |
| 11215 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8) { |
| 11216 | TEST_REQUIRES_ARM_NEON; |
| 11217 | for (uint32_t n = 9; n < 16; n++) { |
| 11218 | for (size_t k = 1; k <= 40; k += 9) { |
| 11219 | GemmMicrokernelTester() |
| 11220 | .mr(4) |
| 11221 | .nr(8) |
| 11222 | .kr(2) |
| 11223 | .sr(1) |
| 11224 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 11225 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11226 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11227 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11228 | } |
| 11229 | } |
| 11230 | } |
| 11231 | |
| 11232 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_strided_cn) { |
| 11233 | TEST_REQUIRES_ARM_NEON; |
| 11234 | for (uint32_t n = 9; n < 16; n++) { |
| 11235 | for (size_t k = 1; k <= 40; k += 9) { |
| 11236 | GemmMicrokernelTester() |
| 11237 | .mr(4) |
| 11238 | .nr(8) |
| 11239 | .kr(2) |
| 11240 | .sr(1) |
| 11241 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 11242 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11243 | .k(k) |
| 11244 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11245 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11246 | } |
| 11247 | } |
| 11248 | } |
| 11249 | |
| 11250 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_strided_a) { |
| 11251 | TEST_REQUIRES_ARM_NEON; |
| 11252 | for (uint32_t n = 9; n < 16; n++) { |
| 11253 | for (size_t k = 1; k <= 40; k += 9) { |
| 11254 | GemmMicrokernelTester() |
| 11255 | .mr(4) |
| 11256 | .nr(8) |
| 11257 | .kr(2) |
| 11258 | .sr(1) |
| 11259 | .m(4) |
| 11260 | .n(n) |
| 11261 | .k(k) |
| 11262 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11263 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11264 | } |
| 11265 | } |
| 11266 | } |
| 11267 | |
| 11268 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_subtile) { |
| 11269 | TEST_REQUIRES_ARM_NEON; |
| 11270 | for (uint32_t n = 9; n < 16; n++) { |
| 11271 | for (size_t k = 1; k <= 40; k += 9) { |
| 11272 | for (uint32_t m = 1; m <= 4; m++) { |
| 11273 | GemmMicrokernelTester() |
| 11274 | .mr(4) |
| 11275 | .nr(8) |
| 11276 | .kr(2) |
| 11277 | .sr(1) |
| 11278 | .m(m) |
| 11279 | .n(n) |
| 11280 | .k(k) |
| 11281 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11282 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11283 | } |
| 11284 | } |
| 11285 | } |
| 11286 | } |
| 11287 | |
| 11288 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8) { |
| 11289 | TEST_REQUIRES_ARM_NEON; |
| 11290 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 11291 | for (size_t k = 1; k <= 40; k += 9) { |
| 11292 | GemmMicrokernelTester() |
| 11293 | .mr(4) |
| 11294 | .nr(8) |
| 11295 | .kr(2) |
| 11296 | .sr(1) |
| 11297 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 11298 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11299 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11300 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11301 | } |
| 11302 | } |
| 11303 | } |
| 11304 | |
| 11305 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_strided_cn) { |
| 11306 | TEST_REQUIRES_ARM_NEON; |
| 11307 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 11308 | for (size_t k = 1; k <= 40; k += 9) { |
| 11309 | GemmMicrokernelTester() |
| 11310 | .mr(4) |
| 11311 | .nr(8) |
| 11312 | .kr(2) |
| 11313 | .sr(1) |
| 11314 | .m(4) |
| 11315 | .n(n) |
| 11316 | .k(k) |
| 11317 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11318 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11319 | } |
| 11320 | } |
| 11321 | } |
| 11322 | |
| 11323 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_strided_a) { |
| 11324 | TEST_REQUIRES_ARM_NEON; |
| 11325 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 11326 | for (size_t k = 1; k <= 40; k += 9) { |
| 11327 | GemmMicrokernelTester() |
| 11328 | .mr(4) |
| 11329 | .nr(8) |
| 11330 | .kr(2) |
| 11331 | .sr(1) |
| 11332 | .m(4) |
| 11333 | .n(n) |
| 11334 | .k(k) |
| 11335 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11336 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11337 | } |
| 11338 | } |
| 11339 | } |
| 11340 | |
| 11341 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_subtile) { |
| 11342 | TEST_REQUIRES_ARM_NEON; |
| 11343 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 11344 | for (size_t k = 1; k <= 40; k += 9) { |
| 11345 | for (uint32_t m = 1; m <= 4; m++) { |
| 11346 | GemmMicrokernelTester() |
| 11347 | .mr(4) |
| 11348 | .nr(8) |
| 11349 | .kr(2) |
| 11350 | .sr(1) |
| 11351 | .m(m) |
| 11352 | .n(n) |
| 11353 | .k(k) |
| 11354 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11355 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11356 | } |
| 11357 | } |
| 11358 | } |
| 11359 | } |
| 11360 | |
| 11361 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cm_subtile) { |
| 11362 | TEST_REQUIRES_ARM_NEON; |
| 11363 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11364 | for (uint32_t n = 1; n <= 8; n++) { |
| 11365 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11366 | GemmMicrokernelTester() |
| 11367 | .mr(4) |
| 11368 | .nr(8) |
| 11369 | .kr(2) |
| 11370 | .sr(1) |
| 11371 | .m(m) |
| 11372 | .n(n) |
| 11373 | .k(k) |
| 11374 | .cm_stride(11) |
| 11375 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11376 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11377 | } |
| 11378 | } |
| 11379 | } |
| 11380 | } |
| 11381 | |
| 11382 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, qmin) { |
| 11383 | TEST_REQUIRES_ARM_NEON; |
| 11384 | GemmMicrokernelTester() |
| 11385 | .mr(4) |
| 11386 | .nr(8) |
| 11387 | .kr(2) |
| 11388 | .sr(1) |
| 11389 | .m(4) |
| 11390 | .n(8) |
| 11391 | .k(8) |
| 11392 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11393 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11394 | } |
| 11395 | |
| 11396 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, qmax) { |
| 11397 | TEST_REQUIRES_ARM_NEON; |
| 11398 | GemmMicrokernelTester() |
| 11399 | .mr(4) |
| 11400 | .nr(8) |
| 11401 | .kr(2) |
| 11402 | .sr(1) |
| 11403 | .m(4) |
| 11404 | .n(8) |
| 11405 | .k(8) |
| 11406 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11407 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11408 | } |
| 11409 | |
| 11410 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cm) { |
| 11411 | TEST_REQUIRES_ARM_NEON; |
| 11412 | GemmMicrokernelTester() |
| 11413 | .mr(4) |
| 11414 | .nr(8) |
| 11415 | .kr(2) |
| 11416 | .sr(1) |
| 11417 | .m(4) |
| 11418 | .n(8) |
| 11419 | .k(8) |
| 11420 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11421 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11422 | } |
| 11423 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 11424 | |
| 11425 | |
| 11426 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 11427 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_eq_8) { |
| 11428 | TEST_REQUIRES_ARM_NEON; |
| 11429 | GemmMicrokernelTester() |
| 11430 | .mr(3) |
| 11431 | .nr(16) |
| 11432 | .kr(2) |
| 11433 | .sr(1) |
| 11434 | .m(3) |
| 11435 | .n(16) |
| 11436 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11437 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11438 | } |
| 11439 | |
| 11440 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, strided_cn) { |
| 11441 | TEST_REQUIRES_ARM_NEON; |
| 11442 | GemmMicrokernelTester() |
| 11443 | .mr(3) |
| 11444 | .nr(16) |
| 11445 | .kr(2) |
| 11446 | .sr(1) |
| 11447 | .m(3) |
| 11448 | .n(16) |
| 11449 | .k(8) |
| 11450 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11451 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11452 | } |
| 11453 | |
| 11454 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_eq_8_strided_a) { |
| 11455 | TEST_REQUIRES_ARM_NEON; |
| 11456 | GemmMicrokernelTester() |
| 11457 | .mr(3) |
| 11458 | .nr(16) |
| 11459 | .kr(2) |
| 11460 | .sr(1) |
| 11461 | .m(3) |
| 11462 | .n(16) |
| 11463 | .k(8) |
| 11464 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11465 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11466 | } |
| 11467 | |
| 11468 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_eq_8_subtile) { |
| 11469 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11470 | for (uint32_t n = 1; n <= 16; n++) { |
| 11471 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11472 | GemmMicrokernelTester() |
| 11473 | .mr(3) |
| 11474 | .nr(16) |
| 11475 | .kr(2) |
| 11476 | .sr(1) |
| 11477 | .m(m) |
| 11478 | .n(n) |
| 11479 | .k(8) |
| 11480 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11481 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11482 | } |
| 11483 | } |
| 11484 | } |
| 11485 | |
| 11486 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_eq_8_subtile_m) { |
| 11487 | TEST_REQUIRES_ARM_NEON; |
| 11488 | for (uint32_t m = 1; m <= 3; m++) { |
| 11489 | GemmMicrokernelTester() |
| 11490 | .mr(3) |
| 11491 | .nr(16) |
| 11492 | .kr(2) |
| 11493 | .sr(1) |
| 11494 | .m(m) |
| 11495 | .n(16) |
| 11496 | .k(8) |
| 11497 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11498 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11499 | } |
| 11500 | } |
| 11501 | |
| 11502 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_eq_8_subtile_n) { |
| 11503 | TEST_REQUIRES_ARM_NEON; |
| 11504 | for (uint32_t n = 1; n <= 16; n++) { |
| 11505 | GemmMicrokernelTester() |
| 11506 | .mr(3) |
| 11507 | .nr(16) |
| 11508 | .kr(2) |
| 11509 | .sr(1) |
| 11510 | .m(3) |
| 11511 | .n(n) |
| 11512 | .k(8) |
| 11513 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11514 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11515 | } |
| 11516 | } |
| 11517 | |
| 11518 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_lt_8) { |
| 11519 | TEST_REQUIRES_ARM_NEON; |
| 11520 | for (size_t k = 1; k < 8; k++) { |
| 11521 | GemmMicrokernelTester() |
| 11522 | .mr(3) |
| 11523 | .nr(16) |
| 11524 | .kr(2) |
| 11525 | .sr(1) |
| 11526 | .m(3) |
| 11527 | .n(16) |
| 11528 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11529 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11530 | } |
| 11531 | } |
| 11532 | |
| 11533 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_lt_8_strided_a) { |
| 11534 | TEST_REQUIRES_ARM_NEON; |
| 11535 | for (size_t k = 1; k < 8; k++) { |
| 11536 | GemmMicrokernelTester() |
| 11537 | .mr(3) |
| 11538 | .nr(16) |
| 11539 | .kr(2) |
| 11540 | .sr(1) |
| 11541 | .m(3) |
| 11542 | .n(16) |
| 11543 | .k(k) |
| 11544 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11545 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11546 | } |
| 11547 | } |
| 11548 | |
| 11549 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_lt_8_subtile) { |
| 11550 | TEST_REQUIRES_ARM_NEON; |
| 11551 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11552 | for (uint32_t n = 1; n <= 16; n++) { |
| 11553 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11554 | GemmMicrokernelTester() |
| 11555 | .mr(3) |
| 11556 | .nr(16) |
| 11557 | .kr(2) |
| 11558 | .sr(1) |
| 11559 | .m(m) |
| 11560 | .n(n) |
| 11561 | .k(k) |
| 11562 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11563 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11564 | } |
| 11565 | } |
| 11566 | } |
| 11567 | } |
| 11568 | |
| 11569 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_gt_8) { |
| 11570 | TEST_REQUIRES_ARM_NEON; |
| 11571 | for (size_t k = 9; k < 16; k++) { |
| 11572 | GemmMicrokernelTester() |
| 11573 | .mr(3) |
| 11574 | .nr(16) |
| 11575 | .kr(2) |
| 11576 | .sr(1) |
| 11577 | .m(3) |
| 11578 | .n(16) |
| 11579 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11580 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11581 | } |
| 11582 | } |
| 11583 | |
| 11584 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_gt_8_strided_a) { |
| 11585 | TEST_REQUIRES_ARM_NEON; |
| 11586 | for (size_t k = 9; k < 16; k++) { |
| 11587 | GemmMicrokernelTester() |
| 11588 | .mr(3) |
| 11589 | .nr(16) |
| 11590 | .kr(2) |
| 11591 | .sr(1) |
| 11592 | .m(3) |
| 11593 | .n(16) |
| 11594 | .k(k) |
| 11595 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11596 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11597 | } |
| 11598 | } |
| 11599 | |
| 11600 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_gt_8_subtile) { |
| 11601 | TEST_REQUIRES_ARM_NEON; |
| 11602 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11603 | for (uint32_t n = 1; n <= 16; n++) { |
| 11604 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11605 | GemmMicrokernelTester() |
| 11606 | .mr(3) |
| 11607 | .nr(16) |
| 11608 | .kr(2) |
| 11609 | .sr(1) |
| 11610 | .m(m) |
| 11611 | .n(n) |
| 11612 | .k(k) |
| 11613 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11614 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11615 | } |
| 11616 | } |
| 11617 | } |
| 11618 | } |
| 11619 | |
| 11620 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_div_8) { |
| 11621 | TEST_REQUIRES_ARM_NEON; |
| 11622 | for (size_t k = 16; k <= 80; k += 8) { |
| 11623 | GemmMicrokernelTester() |
| 11624 | .mr(3) |
| 11625 | .nr(16) |
| 11626 | .kr(2) |
| 11627 | .sr(1) |
| 11628 | .m(3) |
| 11629 | .n(16) |
| 11630 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11631 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11632 | } |
| 11633 | } |
| 11634 | |
| 11635 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_div_8_strided_a) { |
| 11636 | TEST_REQUIRES_ARM_NEON; |
| 11637 | for (size_t k = 16; k <= 80; k += 8) { |
| 11638 | GemmMicrokernelTester() |
| 11639 | .mr(3) |
| 11640 | .nr(16) |
| 11641 | .kr(2) |
| 11642 | .sr(1) |
| 11643 | .m(3) |
| 11644 | .n(16) |
| 11645 | .k(k) |
| 11646 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11647 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11648 | } |
| 11649 | } |
| 11650 | |
| 11651 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, k_div_8_subtile) { |
| 11652 | TEST_REQUIRES_ARM_NEON; |
| 11653 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11654 | for (uint32_t n = 1; n <= 16; n++) { |
| 11655 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11656 | GemmMicrokernelTester() |
| 11657 | .mr(3) |
| 11658 | .nr(16) |
| 11659 | .kr(2) |
| 11660 | .sr(1) |
| 11661 | .m(m) |
| 11662 | .n(n) |
| 11663 | .k(k) |
| 11664 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11665 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11666 | } |
| 11667 | } |
| 11668 | } |
| 11669 | } |
| 11670 | |
| 11671 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_gt_16) { |
| 11672 | TEST_REQUIRES_ARM_NEON; |
| 11673 | for (uint32_t n = 17; n < 32; n++) { |
| 11674 | for (size_t k = 1; k <= 40; k += 9) { |
| 11675 | GemmMicrokernelTester() |
| 11676 | .mr(3) |
| 11677 | .nr(16) |
| 11678 | .kr(2) |
| 11679 | .sr(1) |
| 11680 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 11681 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11682 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11683 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11684 | } |
| 11685 | } |
| 11686 | } |
| 11687 | |
| 11688 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_gt_16_strided_cn) { |
| 11689 | TEST_REQUIRES_ARM_NEON; |
| 11690 | for (uint32_t n = 17; n < 32; n++) { |
| 11691 | for (size_t k = 1; k <= 40; k += 9) { |
| 11692 | GemmMicrokernelTester() |
| 11693 | .mr(3) |
| 11694 | .nr(16) |
| 11695 | .kr(2) |
| 11696 | .sr(1) |
| 11697 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 11698 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11699 | .k(k) |
| 11700 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11701 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11702 | } |
| 11703 | } |
| 11704 | } |
| 11705 | |
| 11706 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_gt_16_strided_a) { |
| 11707 | TEST_REQUIRES_ARM_NEON; |
| 11708 | for (uint32_t n = 17; n < 32; n++) { |
| 11709 | for (size_t k = 1; k <= 40; k += 9) { |
| 11710 | GemmMicrokernelTester() |
| 11711 | .mr(3) |
| 11712 | .nr(16) |
| 11713 | .kr(2) |
| 11714 | .sr(1) |
| 11715 | .m(3) |
| 11716 | .n(n) |
| 11717 | .k(k) |
| 11718 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11719 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11720 | } |
| 11721 | } |
| 11722 | } |
| 11723 | |
| 11724 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_gt_16_subtile) { |
| 11725 | TEST_REQUIRES_ARM_NEON; |
| 11726 | for (uint32_t n = 17; n < 32; n++) { |
| 11727 | for (size_t k = 1; k <= 40; k += 9) { |
| 11728 | for (uint32_t m = 1; m <= 3; m++) { |
| 11729 | GemmMicrokernelTester() |
| 11730 | .mr(3) |
| 11731 | .nr(16) |
| 11732 | .kr(2) |
| 11733 | .sr(1) |
| 11734 | .m(m) |
| 11735 | .n(n) |
| 11736 | .k(k) |
| 11737 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11738 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11739 | } |
| 11740 | } |
| 11741 | } |
| 11742 | } |
| 11743 | |
| 11744 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_div_16) { |
| 11745 | TEST_REQUIRES_ARM_NEON; |
| 11746 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 11747 | for (size_t k = 1; k <= 40; k += 9) { |
| 11748 | GemmMicrokernelTester() |
| 11749 | .mr(3) |
| 11750 | .nr(16) |
| 11751 | .kr(2) |
| 11752 | .sr(1) |
| 11753 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 11754 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11755 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11756 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11757 | } |
| 11758 | } |
| 11759 | } |
| 11760 | |
| 11761 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_div_16_strided_cn) { |
| 11762 | TEST_REQUIRES_ARM_NEON; |
| 11763 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 11764 | for (size_t k = 1; k <= 40; k += 9) { |
| 11765 | GemmMicrokernelTester() |
| 11766 | .mr(3) |
| 11767 | .nr(16) |
| 11768 | .kr(2) |
| 11769 | .sr(1) |
| 11770 | .m(3) |
| 11771 | .n(n) |
| 11772 | .k(k) |
| 11773 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11774 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11775 | } |
| 11776 | } |
| 11777 | } |
| 11778 | |
| 11779 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_div_16_strided_a) { |
| 11780 | TEST_REQUIRES_ARM_NEON; |
| 11781 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 11782 | for (size_t k = 1; k <= 40; k += 9) { |
| 11783 | GemmMicrokernelTester() |
| 11784 | .mr(3) |
| 11785 | .nr(16) |
| 11786 | .kr(2) |
| 11787 | .sr(1) |
| 11788 | .m(3) |
| 11789 | .n(n) |
| 11790 | .k(k) |
| 11791 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11792 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11793 | } |
| 11794 | } |
| 11795 | } |
| 11796 | |
| 11797 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, n_div_16_subtile) { |
| 11798 | TEST_REQUIRES_ARM_NEON; |
| 11799 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 11800 | for (size_t k = 1; k <= 40; k += 9) { |
| 11801 | for (uint32_t m = 1; m <= 3; m++) { |
| 11802 | GemmMicrokernelTester() |
| 11803 | .mr(3) |
| 11804 | .nr(16) |
| 11805 | .kr(2) |
| 11806 | .sr(1) |
| 11807 | .m(m) |
| 11808 | .n(n) |
| 11809 | .k(k) |
| 11810 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11811 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11812 | } |
| 11813 | } |
| 11814 | } |
| 11815 | } |
| 11816 | |
| 11817 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, strided_cm_subtile) { |
| 11818 | TEST_REQUIRES_ARM_NEON; |
| 11819 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11820 | for (uint32_t n = 1; n <= 16; n++) { |
| 11821 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11822 | GemmMicrokernelTester() |
| 11823 | .mr(3) |
| 11824 | .nr(16) |
| 11825 | .kr(2) |
| 11826 | .sr(1) |
| 11827 | .m(m) |
| 11828 | .n(n) |
| 11829 | .k(k) |
| 11830 | .cm_stride(19) |
| 11831 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11832 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11833 | } |
| 11834 | } |
| 11835 | } |
| 11836 | } |
| 11837 | |
| 11838 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, qmin) { |
| 11839 | TEST_REQUIRES_ARM_NEON; |
| 11840 | GemmMicrokernelTester() |
| 11841 | .mr(3) |
| 11842 | .nr(16) |
| 11843 | .kr(2) |
| 11844 | .sr(1) |
| 11845 | .m(3) |
| 11846 | .n(16) |
| 11847 | .k(8) |
| 11848 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11849 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11850 | } |
| 11851 | |
| 11852 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, qmax) { |
| 11853 | TEST_REQUIRES_ARM_NEON; |
| 11854 | GemmMicrokernelTester() |
| 11855 | .mr(3) |
| 11856 | .nr(16) |
| 11857 | .kr(2) |
| 11858 | .sr(1) |
| 11859 | .m(3) |
| 11860 | .n(16) |
| 11861 | .k(8) |
| 11862 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11863 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11864 | } |
| 11865 | |
| 11866 | TEST(QS8_GEMM_MINMAX_RNDNU_3X16C2__NEON_MULL_DUP, strided_cm) { |
| 11867 | TEST_REQUIRES_ARM_NEON; |
| 11868 | GemmMicrokernelTester() |
| 11869 | .mr(3) |
| 11870 | .nr(16) |
| 11871 | .kr(2) |
| 11872 | .sr(1) |
| 11873 | .m(3) |
| 11874 | .n(16) |
| 11875 | .k(8) |
| 11876 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11877 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11878 | } |
| 11879 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 11880 | |
| 11881 | |
| 11882 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 11883 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8) { |
| 11884 | TEST_REQUIRES_ARM_NEON; |
| 11885 | GemmMicrokernelTester() |
| 11886 | .mr(4) |
| 11887 | .nr(16) |
| 11888 | .kr(2) |
| 11889 | .sr(1) |
| 11890 | .m(4) |
| 11891 | .n(16) |
| 11892 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11893 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11894 | } |
| 11895 | |
| 11896 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, strided_cn) { |
| 11897 | TEST_REQUIRES_ARM_NEON; |
| 11898 | GemmMicrokernelTester() |
| 11899 | .mr(4) |
| 11900 | .nr(16) |
| 11901 | .kr(2) |
| 11902 | .sr(1) |
| 11903 | .m(4) |
| 11904 | .n(16) |
| 11905 | .k(8) |
| 11906 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11907 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11908 | } |
| 11909 | |
| 11910 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8_strided_a) { |
| 11911 | TEST_REQUIRES_ARM_NEON; |
| 11912 | GemmMicrokernelTester() |
| 11913 | .mr(4) |
| 11914 | .nr(16) |
| 11915 | .kr(2) |
| 11916 | .sr(1) |
| 11917 | .m(4) |
| 11918 | .n(16) |
| 11919 | .k(8) |
| 11920 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11921 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11922 | } |
| 11923 | |
| 11924 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8_subtile) { |
| 11925 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 11926 | for (uint32_t n = 1; n <= 16; n++) { |
| 11927 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11928 | GemmMicrokernelTester() |
| 11929 | .mr(4) |
| 11930 | .nr(16) |
| 11931 | .kr(2) |
| 11932 | .sr(1) |
| 11933 | .m(m) |
| 11934 | .n(n) |
| 11935 | .k(8) |
| 11936 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11937 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11938 | } |
| 11939 | } |
| 11940 | } |
| 11941 | |
| 11942 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8_subtile_m) { |
| 11943 | TEST_REQUIRES_ARM_NEON; |
| 11944 | for (uint32_t m = 1; m <= 4; m++) { |
| 11945 | GemmMicrokernelTester() |
| 11946 | .mr(4) |
| 11947 | .nr(16) |
| 11948 | .kr(2) |
| 11949 | .sr(1) |
| 11950 | .m(m) |
| 11951 | .n(16) |
| 11952 | .k(8) |
| 11953 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11954 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11955 | } |
| 11956 | } |
| 11957 | |
| 11958 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_eq_8_subtile_n) { |
| 11959 | TEST_REQUIRES_ARM_NEON; |
| 11960 | for (uint32_t n = 1; n <= 16; n++) { |
| 11961 | GemmMicrokernelTester() |
| 11962 | .mr(4) |
| 11963 | .nr(16) |
| 11964 | .kr(2) |
| 11965 | .sr(1) |
| 11966 | .m(4) |
| 11967 | .n(n) |
| 11968 | .k(8) |
| 11969 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11970 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11971 | } |
| 11972 | } |
| 11973 | |
| 11974 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_lt_8) { |
| 11975 | TEST_REQUIRES_ARM_NEON; |
| 11976 | for (size_t k = 1; k < 8; k++) { |
| 11977 | GemmMicrokernelTester() |
| 11978 | .mr(4) |
| 11979 | .nr(16) |
| 11980 | .kr(2) |
| 11981 | .sr(1) |
| 11982 | .m(4) |
| 11983 | .n(16) |
| 11984 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 11985 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 11986 | } |
| 11987 | } |
| 11988 | |
| 11989 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_lt_8_strided_a) { |
| 11990 | TEST_REQUIRES_ARM_NEON; |
| 11991 | for (size_t k = 1; k < 8; k++) { |
| 11992 | GemmMicrokernelTester() |
| 11993 | .mr(4) |
| 11994 | .nr(16) |
| 11995 | .kr(2) |
| 11996 | .sr(1) |
| 11997 | .m(4) |
| 11998 | .n(16) |
| 11999 | .k(k) |
| 12000 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12001 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12002 | } |
| 12003 | } |
| 12004 | |
| 12005 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_lt_8_subtile) { |
| 12006 | TEST_REQUIRES_ARM_NEON; |
| 12007 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12008 | for (uint32_t n = 1; n <= 16; n++) { |
| 12009 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12010 | GemmMicrokernelTester() |
| 12011 | .mr(4) |
| 12012 | .nr(16) |
| 12013 | .kr(2) |
| 12014 | .sr(1) |
| 12015 | .m(m) |
| 12016 | .n(n) |
| 12017 | .k(k) |
| 12018 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12019 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12020 | } |
| 12021 | } |
| 12022 | } |
| 12023 | } |
| 12024 | |
| 12025 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_gt_8) { |
| 12026 | TEST_REQUIRES_ARM_NEON; |
| 12027 | for (size_t k = 9; k < 16; k++) { |
| 12028 | GemmMicrokernelTester() |
| 12029 | .mr(4) |
| 12030 | .nr(16) |
| 12031 | .kr(2) |
| 12032 | .sr(1) |
| 12033 | .m(4) |
| 12034 | .n(16) |
| 12035 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12036 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12037 | } |
| 12038 | } |
| 12039 | |
| 12040 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_gt_8_strided_a) { |
| 12041 | TEST_REQUIRES_ARM_NEON; |
| 12042 | for (size_t k = 9; k < 16; k++) { |
| 12043 | GemmMicrokernelTester() |
| 12044 | .mr(4) |
| 12045 | .nr(16) |
| 12046 | .kr(2) |
| 12047 | .sr(1) |
| 12048 | .m(4) |
| 12049 | .n(16) |
| 12050 | .k(k) |
| 12051 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12052 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12053 | } |
| 12054 | } |
| 12055 | |
| 12056 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_gt_8_subtile) { |
| 12057 | TEST_REQUIRES_ARM_NEON; |
| 12058 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12059 | for (uint32_t n = 1; n <= 16; n++) { |
| 12060 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12061 | GemmMicrokernelTester() |
| 12062 | .mr(4) |
| 12063 | .nr(16) |
| 12064 | .kr(2) |
| 12065 | .sr(1) |
| 12066 | .m(m) |
| 12067 | .n(n) |
| 12068 | .k(k) |
| 12069 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12070 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12071 | } |
| 12072 | } |
| 12073 | } |
| 12074 | } |
| 12075 | |
| 12076 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_div_8) { |
| 12077 | TEST_REQUIRES_ARM_NEON; |
| 12078 | for (size_t k = 16; k <= 80; k += 8) { |
| 12079 | GemmMicrokernelTester() |
| 12080 | .mr(4) |
| 12081 | .nr(16) |
| 12082 | .kr(2) |
| 12083 | .sr(1) |
| 12084 | .m(4) |
| 12085 | .n(16) |
| 12086 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12087 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12088 | } |
| 12089 | } |
| 12090 | |
| 12091 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_div_8_strided_a) { |
| 12092 | TEST_REQUIRES_ARM_NEON; |
| 12093 | for (size_t k = 16; k <= 80; k += 8) { |
| 12094 | GemmMicrokernelTester() |
| 12095 | .mr(4) |
| 12096 | .nr(16) |
| 12097 | .kr(2) |
| 12098 | .sr(1) |
| 12099 | .m(4) |
| 12100 | .n(16) |
| 12101 | .k(k) |
| 12102 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12103 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12104 | } |
| 12105 | } |
| 12106 | |
| 12107 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, k_div_8_subtile) { |
| 12108 | TEST_REQUIRES_ARM_NEON; |
| 12109 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12110 | for (uint32_t n = 1; n <= 16; n++) { |
| 12111 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12112 | GemmMicrokernelTester() |
| 12113 | .mr(4) |
| 12114 | .nr(16) |
| 12115 | .kr(2) |
| 12116 | .sr(1) |
| 12117 | .m(m) |
| 12118 | .n(n) |
| 12119 | .k(k) |
| 12120 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12121 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12122 | } |
| 12123 | } |
| 12124 | } |
| 12125 | } |
| 12126 | |
| 12127 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16) { |
| 12128 | TEST_REQUIRES_ARM_NEON; |
| 12129 | for (uint32_t n = 17; n < 32; n++) { |
| 12130 | for (size_t k = 1; k <= 40; k += 9) { |
| 12131 | GemmMicrokernelTester() |
| 12132 | .mr(4) |
| 12133 | .nr(16) |
| 12134 | .kr(2) |
| 12135 | .sr(1) |
| 12136 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 12137 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12138 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12139 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12140 | } |
| 12141 | } |
| 12142 | } |
| 12143 | |
| 12144 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16_strided_cn) { |
| 12145 | TEST_REQUIRES_ARM_NEON; |
| 12146 | for (uint32_t n = 17; n < 32; n++) { |
| 12147 | for (size_t k = 1; k <= 40; k += 9) { |
| 12148 | GemmMicrokernelTester() |
| 12149 | .mr(4) |
| 12150 | .nr(16) |
| 12151 | .kr(2) |
| 12152 | .sr(1) |
| 12153 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 12154 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12155 | .k(k) |
| 12156 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12157 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12158 | } |
| 12159 | } |
| 12160 | } |
| 12161 | |
| 12162 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16_strided_a) { |
| 12163 | TEST_REQUIRES_ARM_NEON; |
| 12164 | for (uint32_t n = 17; n < 32; n++) { |
| 12165 | for (size_t k = 1; k <= 40; k += 9) { |
| 12166 | GemmMicrokernelTester() |
| 12167 | .mr(4) |
| 12168 | .nr(16) |
| 12169 | .kr(2) |
| 12170 | .sr(1) |
| 12171 | .m(4) |
| 12172 | .n(n) |
| 12173 | .k(k) |
| 12174 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12175 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12176 | } |
| 12177 | } |
| 12178 | } |
| 12179 | |
| 12180 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_gt_16_subtile) { |
| 12181 | TEST_REQUIRES_ARM_NEON; |
| 12182 | for (uint32_t n = 17; n < 32; n++) { |
| 12183 | for (size_t k = 1; k <= 40; k += 9) { |
| 12184 | for (uint32_t m = 1; m <= 4; m++) { |
| 12185 | GemmMicrokernelTester() |
| 12186 | .mr(4) |
| 12187 | .nr(16) |
| 12188 | .kr(2) |
| 12189 | .sr(1) |
| 12190 | .m(m) |
| 12191 | .n(n) |
| 12192 | .k(k) |
| 12193 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12194 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12195 | } |
| 12196 | } |
| 12197 | } |
| 12198 | } |
| 12199 | |
| 12200 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16) { |
| 12201 | TEST_REQUIRES_ARM_NEON; |
| 12202 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 12203 | for (size_t k = 1; k <= 40; k += 9) { |
| 12204 | GemmMicrokernelTester() |
| 12205 | .mr(4) |
| 12206 | .nr(16) |
| 12207 | .kr(2) |
| 12208 | .sr(1) |
| 12209 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 12210 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12211 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12212 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12213 | } |
| 12214 | } |
| 12215 | } |
| 12216 | |
| 12217 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16_strided_cn) { |
| 12218 | TEST_REQUIRES_ARM_NEON; |
| 12219 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 12220 | for (size_t k = 1; k <= 40; k += 9) { |
| 12221 | GemmMicrokernelTester() |
| 12222 | .mr(4) |
| 12223 | .nr(16) |
| 12224 | .kr(2) |
| 12225 | .sr(1) |
| 12226 | .m(4) |
| 12227 | .n(n) |
| 12228 | .k(k) |
| 12229 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12230 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12231 | } |
| 12232 | } |
| 12233 | } |
| 12234 | |
| 12235 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16_strided_a) { |
| 12236 | TEST_REQUIRES_ARM_NEON; |
| 12237 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 12238 | for (size_t k = 1; k <= 40; k += 9) { |
| 12239 | GemmMicrokernelTester() |
| 12240 | .mr(4) |
| 12241 | .nr(16) |
| 12242 | .kr(2) |
| 12243 | .sr(1) |
| 12244 | .m(4) |
| 12245 | .n(n) |
| 12246 | .k(k) |
| 12247 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12248 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12249 | } |
| 12250 | } |
| 12251 | } |
| 12252 | |
| 12253 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, n_div_16_subtile) { |
| 12254 | TEST_REQUIRES_ARM_NEON; |
| 12255 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 12256 | for (size_t k = 1; k <= 40; k += 9) { |
| 12257 | for (uint32_t m = 1; m <= 4; m++) { |
| 12258 | GemmMicrokernelTester() |
| 12259 | .mr(4) |
| 12260 | .nr(16) |
| 12261 | .kr(2) |
| 12262 | .sr(1) |
| 12263 | .m(m) |
| 12264 | .n(n) |
| 12265 | .k(k) |
| 12266 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12267 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12268 | } |
| 12269 | } |
| 12270 | } |
| 12271 | } |
| 12272 | |
| 12273 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, strided_cm_subtile) { |
| 12274 | TEST_REQUIRES_ARM_NEON; |
| 12275 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12276 | for (uint32_t n = 1; n <= 16; n++) { |
| 12277 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12278 | GemmMicrokernelTester() |
| 12279 | .mr(4) |
| 12280 | .nr(16) |
| 12281 | .kr(2) |
| 12282 | .sr(1) |
| 12283 | .m(m) |
| 12284 | .n(n) |
| 12285 | .k(k) |
| 12286 | .cm_stride(19) |
| 12287 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12288 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12289 | } |
| 12290 | } |
| 12291 | } |
| 12292 | } |
| 12293 | |
| 12294 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, qmin) { |
| 12295 | TEST_REQUIRES_ARM_NEON; |
| 12296 | GemmMicrokernelTester() |
| 12297 | .mr(4) |
| 12298 | .nr(16) |
| 12299 | .kr(2) |
| 12300 | .sr(1) |
| 12301 | .m(4) |
| 12302 | .n(16) |
| 12303 | .k(8) |
| 12304 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12305 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12306 | } |
| 12307 | |
| 12308 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, qmax) { |
| 12309 | TEST_REQUIRES_ARM_NEON; |
| 12310 | GemmMicrokernelTester() |
| 12311 | .mr(4) |
| 12312 | .nr(16) |
| 12313 | .kr(2) |
| 12314 | .sr(1) |
| 12315 | .m(4) |
| 12316 | .n(16) |
| 12317 | .k(8) |
| 12318 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12319 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12320 | } |
| 12321 | |
| 12322 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C2__NEON_MULL_DUP, strided_cm) { |
| 12323 | TEST_REQUIRES_ARM_NEON; |
| 12324 | GemmMicrokernelTester() |
| 12325 | .mr(4) |
| 12326 | .nr(16) |
| 12327 | .kr(2) |
| 12328 | .sr(1) |
| 12329 | .m(4) |
| 12330 | .n(16) |
| 12331 | .k(8) |
| 12332 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12333 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12334 | } |
| 12335 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 12336 | |
| 12337 | |
| 12338 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 12339 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16) { |
| 12340 | TEST_REQUIRES_ARM_NEON; |
| 12341 | GemmMicrokernelTester() |
| 12342 | .mr(2) |
| 12343 | .nr(8) |
| 12344 | .kr(2) |
| 12345 | .sr(1) |
| 12346 | .m(2) |
| 12347 | .n(8) |
| 12348 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12349 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12350 | } |
| 12351 | |
| 12352 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, strided_cn) { |
| 12353 | TEST_REQUIRES_ARM_NEON; |
| 12354 | GemmMicrokernelTester() |
| 12355 | .mr(2) |
| 12356 | .nr(8) |
| 12357 | .kr(2) |
| 12358 | .sr(1) |
| 12359 | .m(2) |
| 12360 | .n(8) |
| 12361 | .k(16) |
| 12362 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12363 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12364 | } |
| 12365 | |
| 12366 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16_strided_a) { |
| 12367 | TEST_REQUIRES_ARM_NEON; |
| 12368 | GemmMicrokernelTester() |
| 12369 | .mr(2) |
| 12370 | .nr(8) |
| 12371 | .kr(2) |
| 12372 | .sr(1) |
| 12373 | .m(2) |
| 12374 | .n(8) |
| 12375 | .k(16) |
| 12376 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12377 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12378 | } |
| 12379 | |
| 12380 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16_subtile) { |
| 12381 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12382 | for (uint32_t n = 1; n <= 8; n++) { |
| 12383 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12384 | GemmMicrokernelTester() |
| 12385 | .mr(2) |
| 12386 | .nr(8) |
| 12387 | .kr(2) |
| 12388 | .sr(1) |
| 12389 | .m(m) |
| 12390 | .n(n) |
| 12391 | .k(16) |
| 12392 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12393 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12394 | } |
| 12395 | } |
| 12396 | } |
| 12397 | |
| 12398 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) { |
| 12399 | TEST_REQUIRES_ARM_NEON; |
| 12400 | for (uint32_t m = 1; m <= 2; m++) { |
| 12401 | GemmMicrokernelTester() |
| 12402 | .mr(2) |
| 12403 | .nr(8) |
| 12404 | .kr(2) |
| 12405 | .sr(1) |
| 12406 | .m(m) |
| 12407 | .n(8) |
| 12408 | .k(16) |
| 12409 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12410 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12411 | } |
| 12412 | } |
| 12413 | |
| 12414 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) { |
| 12415 | TEST_REQUIRES_ARM_NEON; |
| 12416 | for (uint32_t n = 1; n <= 8; n++) { |
| 12417 | GemmMicrokernelTester() |
| 12418 | .mr(2) |
| 12419 | .nr(8) |
| 12420 | .kr(2) |
| 12421 | .sr(1) |
| 12422 | .m(2) |
| 12423 | .n(n) |
| 12424 | .k(16) |
| 12425 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12426 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12427 | } |
| 12428 | } |
| 12429 | |
| 12430 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_lt_16) { |
| 12431 | TEST_REQUIRES_ARM_NEON; |
| 12432 | for (size_t k = 1; k < 16; k++) { |
| 12433 | GemmMicrokernelTester() |
| 12434 | .mr(2) |
| 12435 | .nr(8) |
| 12436 | .kr(2) |
| 12437 | .sr(1) |
| 12438 | .m(2) |
| 12439 | .n(8) |
| 12440 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12441 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12442 | } |
| 12443 | } |
| 12444 | |
| 12445 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_lt_16_strided_a) { |
| 12446 | TEST_REQUIRES_ARM_NEON; |
| 12447 | for (size_t k = 1; k < 16; k++) { |
| 12448 | GemmMicrokernelTester() |
| 12449 | .mr(2) |
| 12450 | .nr(8) |
| 12451 | .kr(2) |
| 12452 | .sr(1) |
| 12453 | .m(2) |
| 12454 | .n(8) |
| 12455 | .k(k) |
| 12456 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12457 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12458 | } |
| 12459 | } |
| 12460 | |
| 12461 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_lt_16_subtile) { |
| 12462 | TEST_REQUIRES_ARM_NEON; |
| 12463 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12464 | for (uint32_t n = 1; n <= 8; n++) { |
| 12465 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12466 | GemmMicrokernelTester() |
| 12467 | .mr(2) |
| 12468 | .nr(8) |
| 12469 | .kr(2) |
| 12470 | .sr(1) |
| 12471 | .m(m) |
| 12472 | .n(n) |
| 12473 | .k(k) |
| 12474 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12475 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12476 | } |
| 12477 | } |
| 12478 | } |
| 12479 | } |
| 12480 | |
| 12481 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_gt_16) { |
| 12482 | TEST_REQUIRES_ARM_NEON; |
| 12483 | for (size_t k = 17; k < 32; k++) { |
| 12484 | GemmMicrokernelTester() |
| 12485 | .mr(2) |
| 12486 | .nr(8) |
| 12487 | .kr(2) |
| 12488 | .sr(1) |
| 12489 | .m(2) |
| 12490 | .n(8) |
| 12491 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12492 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12493 | } |
| 12494 | } |
| 12495 | |
| 12496 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_gt_16_strided_a) { |
| 12497 | TEST_REQUIRES_ARM_NEON; |
| 12498 | for (size_t k = 17; k < 32; k++) { |
| 12499 | GemmMicrokernelTester() |
| 12500 | .mr(2) |
| 12501 | .nr(8) |
| 12502 | .kr(2) |
| 12503 | .sr(1) |
| 12504 | .m(2) |
| 12505 | .n(8) |
| 12506 | .k(k) |
| 12507 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12508 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12509 | } |
| 12510 | } |
| 12511 | |
| 12512 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_gt_16_subtile) { |
| 12513 | TEST_REQUIRES_ARM_NEON; |
| 12514 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12515 | for (uint32_t n = 1; n <= 8; n++) { |
| 12516 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12517 | GemmMicrokernelTester() |
| 12518 | .mr(2) |
| 12519 | .nr(8) |
| 12520 | .kr(2) |
| 12521 | .sr(1) |
| 12522 | .m(m) |
| 12523 | .n(n) |
| 12524 | .k(k) |
| 12525 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12526 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12527 | } |
| 12528 | } |
| 12529 | } |
| 12530 | } |
| 12531 | |
| 12532 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_div_16) { |
| 12533 | TEST_REQUIRES_ARM_NEON; |
| 12534 | for (size_t k = 32; k <= 160; k += 16) { |
| 12535 | GemmMicrokernelTester() |
| 12536 | .mr(2) |
| 12537 | .nr(8) |
| 12538 | .kr(2) |
| 12539 | .sr(1) |
| 12540 | .m(2) |
| 12541 | .n(8) |
| 12542 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12543 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12544 | } |
| 12545 | } |
| 12546 | |
| 12547 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_div_16_strided_a) { |
| 12548 | TEST_REQUIRES_ARM_NEON; |
| 12549 | for (size_t k = 32; k <= 160; k += 16) { |
| 12550 | GemmMicrokernelTester() |
| 12551 | .mr(2) |
| 12552 | .nr(8) |
| 12553 | .kr(2) |
| 12554 | .sr(1) |
| 12555 | .m(2) |
| 12556 | .n(8) |
| 12557 | .k(k) |
| 12558 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12559 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12560 | } |
| 12561 | } |
| 12562 | |
| 12563 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, k_div_16_subtile) { |
| 12564 | TEST_REQUIRES_ARM_NEON; |
| 12565 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12566 | for (uint32_t n = 1; n <= 8; n++) { |
| 12567 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12568 | GemmMicrokernelTester() |
| 12569 | .mr(2) |
| 12570 | .nr(8) |
| 12571 | .kr(2) |
| 12572 | .sr(1) |
| 12573 | .m(m) |
| 12574 | .n(n) |
| 12575 | .k(k) |
| 12576 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12577 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12578 | } |
| 12579 | } |
| 12580 | } |
| 12581 | } |
| 12582 | |
| 12583 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8) { |
| 12584 | TEST_REQUIRES_ARM_NEON; |
| 12585 | for (uint32_t n = 9; n < 16; n++) { |
| 12586 | for (size_t k = 1; k <= 80; k += 17) { |
| 12587 | GemmMicrokernelTester() |
| 12588 | .mr(2) |
| 12589 | .nr(8) |
| 12590 | .kr(2) |
| 12591 | .sr(1) |
| 12592 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 12593 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12594 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12595 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12596 | } |
| 12597 | } |
| 12598 | } |
| 12599 | |
| 12600 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) { |
| 12601 | TEST_REQUIRES_ARM_NEON; |
| 12602 | for (uint32_t n = 9; n < 16; n++) { |
| 12603 | for (size_t k = 1; k <= 80; k += 17) { |
| 12604 | GemmMicrokernelTester() |
| 12605 | .mr(2) |
| 12606 | .nr(8) |
| 12607 | .kr(2) |
| 12608 | .sr(1) |
| 12609 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 12610 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12611 | .k(k) |
| 12612 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12613 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12614 | } |
| 12615 | } |
| 12616 | } |
| 12617 | |
| 12618 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8_strided_a) { |
| 12619 | TEST_REQUIRES_ARM_NEON; |
| 12620 | for (uint32_t n = 9; n < 16; n++) { |
| 12621 | for (size_t k = 1; k <= 80; k += 17) { |
| 12622 | GemmMicrokernelTester() |
| 12623 | .mr(2) |
| 12624 | .nr(8) |
| 12625 | .kr(2) |
| 12626 | .sr(1) |
| 12627 | .m(2) |
| 12628 | .n(n) |
| 12629 | .k(k) |
| 12630 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12631 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12632 | } |
| 12633 | } |
| 12634 | } |
| 12635 | |
| 12636 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_gt_8_subtile) { |
| 12637 | TEST_REQUIRES_ARM_NEON; |
| 12638 | for (uint32_t n = 9; n < 16; n++) { |
| 12639 | for (size_t k = 1; k <= 80; k += 17) { |
| 12640 | for (uint32_t m = 1; m <= 2; m++) { |
| 12641 | GemmMicrokernelTester() |
| 12642 | .mr(2) |
| 12643 | .nr(8) |
| 12644 | .kr(2) |
| 12645 | .sr(1) |
| 12646 | .m(m) |
| 12647 | .n(n) |
| 12648 | .k(k) |
| 12649 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12650 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12651 | } |
| 12652 | } |
| 12653 | } |
| 12654 | } |
| 12655 | |
| 12656 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8) { |
| 12657 | TEST_REQUIRES_ARM_NEON; |
| 12658 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 12659 | for (size_t k = 1; k <= 80; k += 17) { |
| 12660 | GemmMicrokernelTester() |
| 12661 | .mr(2) |
| 12662 | .nr(8) |
| 12663 | .kr(2) |
| 12664 | .sr(1) |
| 12665 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 12666 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12667 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12668 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12669 | } |
| 12670 | } |
| 12671 | } |
| 12672 | |
| 12673 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) { |
| 12674 | TEST_REQUIRES_ARM_NEON; |
| 12675 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 12676 | for (size_t k = 1; k <= 80; k += 17) { |
| 12677 | GemmMicrokernelTester() |
| 12678 | .mr(2) |
| 12679 | .nr(8) |
| 12680 | .kr(2) |
| 12681 | .sr(1) |
| 12682 | .m(2) |
| 12683 | .n(n) |
| 12684 | .k(k) |
| 12685 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12686 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12687 | } |
| 12688 | } |
| 12689 | } |
| 12690 | |
| 12691 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8_strided_a) { |
| 12692 | TEST_REQUIRES_ARM_NEON; |
| 12693 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 12694 | for (size_t k = 1; k <= 80; k += 17) { |
| 12695 | GemmMicrokernelTester() |
| 12696 | .mr(2) |
| 12697 | .nr(8) |
| 12698 | .kr(2) |
| 12699 | .sr(1) |
| 12700 | .m(2) |
| 12701 | .n(n) |
| 12702 | .k(k) |
| 12703 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12704 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12705 | } |
| 12706 | } |
| 12707 | } |
| 12708 | |
| 12709 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, n_div_8_subtile) { |
| 12710 | TEST_REQUIRES_ARM_NEON; |
| 12711 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 12712 | for (size_t k = 1; k <= 80; k += 17) { |
| 12713 | for (uint32_t m = 1; m <= 2; m++) { |
| 12714 | GemmMicrokernelTester() |
| 12715 | .mr(2) |
| 12716 | .nr(8) |
| 12717 | .kr(2) |
| 12718 | .sr(1) |
| 12719 | .m(m) |
| 12720 | .n(n) |
| 12721 | .k(k) |
| 12722 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12723 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12724 | } |
| 12725 | } |
| 12726 | } |
| 12727 | } |
| 12728 | |
| 12729 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, strided_cm_subtile) { |
| 12730 | TEST_REQUIRES_ARM_NEON; |
| 12731 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12732 | for (uint32_t n = 1; n <= 8; n++) { |
| 12733 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12734 | GemmMicrokernelTester() |
| 12735 | .mr(2) |
| 12736 | .nr(8) |
| 12737 | .kr(2) |
| 12738 | .sr(1) |
| 12739 | .m(m) |
| 12740 | .n(n) |
| 12741 | .k(k) |
| 12742 | .cm_stride(11) |
| 12743 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12744 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12745 | } |
| 12746 | } |
| 12747 | } |
| 12748 | } |
| 12749 | |
| 12750 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, qmin) { |
| 12751 | TEST_REQUIRES_ARM_NEON; |
| 12752 | GemmMicrokernelTester() |
| 12753 | .mr(2) |
| 12754 | .nr(8) |
| 12755 | .kr(2) |
| 12756 | .sr(1) |
| 12757 | .m(2) |
| 12758 | .n(8) |
| 12759 | .k(16) |
| 12760 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12761 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12762 | } |
| 12763 | |
| 12764 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, qmax) { |
| 12765 | TEST_REQUIRES_ARM_NEON; |
| 12766 | GemmMicrokernelTester() |
| 12767 | .mr(2) |
| 12768 | .nr(8) |
| 12769 | .kr(2) |
| 12770 | .sr(1) |
| 12771 | .m(2) |
| 12772 | .n(8) |
| 12773 | .k(16) |
| 12774 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12775 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12776 | } |
| 12777 | |
| 12778 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_DUP, strided_cm) { |
| 12779 | TEST_REQUIRES_ARM_NEON; |
| 12780 | GemmMicrokernelTester() |
| 12781 | .mr(2) |
| 12782 | .nr(8) |
| 12783 | .kr(2) |
| 12784 | .sr(1) |
| 12785 | .m(2) |
| 12786 | .n(8) |
| 12787 | .k(16) |
| 12788 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12789 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12790 | } |
| 12791 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 12792 | |
| 12793 | |
| 12794 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 12795 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16) { |
| 12796 | TEST_REQUIRES_ARM_NEON; |
| 12797 | GemmMicrokernelTester() |
| 12798 | .mr(3) |
| 12799 | .nr(8) |
| 12800 | .kr(2) |
| 12801 | .sr(1) |
| 12802 | .m(3) |
| 12803 | .n(8) |
| 12804 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12805 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12806 | } |
| 12807 | |
| 12808 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cn) { |
| 12809 | TEST_REQUIRES_ARM_NEON; |
| 12810 | GemmMicrokernelTester() |
| 12811 | .mr(3) |
| 12812 | .nr(8) |
| 12813 | .kr(2) |
| 12814 | .sr(1) |
| 12815 | .m(3) |
| 12816 | .n(8) |
| 12817 | .k(16) |
| 12818 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12819 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12820 | } |
| 12821 | |
| 12822 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_strided_a) { |
| 12823 | TEST_REQUIRES_ARM_NEON; |
| 12824 | GemmMicrokernelTester() |
| 12825 | .mr(3) |
| 12826 | .nr(8) |
| 12827 | .kr(2) |
| 12828 | .sr(1) |
| 12829 | .m(3) |
| 12830 | .n(8) |
| 12831 | .k(16) |
| 12832 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12833 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12834 | } |
| 12835 | |
| 12836 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile) { |
| 12837 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12838 | for (uint32_t n = 1; n <= 8; n++) { |
| 12839 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12840 | GemmMicrokernelTester() |
| 12841 | .mr(3) |
| 12842 | .nr(8) |
| 12843 | .kr(2) |
| 12844 | .sr(1) |
| 12845 | .m(m) |
| 12846 | .n(n) |
| 12847 | .k(16) |
| 12848 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12849 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12850 | } |
| 12851 | } |
| 12852 | } |
| 12853 | |
| 12854 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) { |
| 12855 | TEST_REQUIRES_ARM_NEON; |
| 12856 | for (uint32_t m = 1; m <= 3; m++) { |
| 12857 | GemmMicrokernelTester() |
| 12858 | .mr(3) |
| 12859 | .nr(8) |
| 12860 | .kr(2) |
| 12861 | .sr(1) |
| 12862 | .m(m) |
| 12863 | .n(8) |
| 12864 | .k(16) |
| 12865 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12866 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12867 | } |
| 12868 | } |
| 12869 | |
| 12870 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) { |
| 12871 | TEST_REQUIRES_ARM_NEON; |
| 12872 | for (uint32_t n = 1; n <= 8; n++) { |
| 12873 | GemmMicrokernelTester() |
| 12874 | .mr(3) |
| 12875 | .nr(8) |
| 12876 | .kr(2) |
| 12877 | .sr(1) |
| 12878 | .m(3) |
| 12879 | .n(n) |
| 12880 | .k(16) |
| 12881 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12882 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12883 | } |
| 12884 | } |
| 12885 | |
| 12886 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_lt_16) { |
| 12887 | TEST_REQUIRES_ARM_NEON; |
| 12888 | for (size_t k = 1; k < 16; k++) { |
| 12889 | GemmMicrokernelTester() |
| 12890 | .mr(3) |
| 12891 | .nr(8) |
| 12892 | .kr(2) |
| 12893 | .sr(1) |
| 12894 | .m(3) |
| 12895 | .n(8) |
| 12896 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12897 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12898 | } |
| 12899 | } |
| 12900 | |
| 12901 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_lt_16_strided_a) { |
| 12902 | TEST_REQUIRES_ARM_NEON; |
| 12903 | for (size_t k = 1; k < 16; k++) { |
| 12904 | GemmMicrokernelTester() |
| 12905 | .mr(3) |
| 12906 | .nr(8) |
| 12907 | .kr(2) |
| 12908 | .sr(1) |
| 12909 | .m(3) |
| 12910 | .n(8) |
| 12911 | .k(k) |
| 12912 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12913 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12914 | } |
| 12915 | } |
| 12916 | |
| 12917 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_lt_16_subtile) { |
| 12918 | TEST_REQUIRES_ARM_NEON; |
| 12919 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12920 | for (uint32_t n = 1; n <= 8; n++) { |
| 12921 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12922 | GemmMicrokernelTester() |
| 12923 | .mr(3) |
| 12924 | .nr(8) |
| 12925 | .kr(2) |
| 12926 | .sr(1) |
| 12927 | .m(m) |
| 12928 | .n(n) |
| 12929 | .k(k) |
| 12930 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12931 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12932 | } |
| 12933 | } |
| 12934 | } |
| 12935 | } |
| 12936 | |
| 12937 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_gt_16) { |
| 12938 | TEST_REQUIRES_ARM_NEON; |
| 12939 | for (size_t k = 17; k < 32; k++) { |
| 12940 | GemmMicrokernelTester() |
| 12941 | .mr(3) |
| 12942 | .nr(8) |
| 12943 | .kr(2) |
| 12944 | .sr(1) |
| 12945 | .m(3) |
| 12946 | .n(8) |
| 12947 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12948 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12949 | } |
| 12950 | } |
| 12951 | |
| 12952 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_gt_16_strided_a) { |
| 12953 | TEST_REQUIRES_ARM_NEON; |
| 12954 | for (size_t k = 17; k < 32; k++) { |
| 12955 | GemmMicrokernelTester() |
| 12956 | .mr(3) |
| 12957 | .nr(8) |
| 12958 | .kr(2) |
| 12959 | .sr(1) |
| 12960 | .m(3) |
| 12961 | .n(8) |
| 12962 | .k(k) |
| 12963 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12964 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12965 | } |
| 12966 | } |
| 12967 | |
| 12968 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_gt_16_subtile) { |
| 12969 | TEST_REQUIRES_ARM_NEON; |
| 12970 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 12971 | for (uint32_t n = 1; n <= 8; n++) { |
| 12972 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12973 | GemmMicrokernelTester() |
| 12974 | .mr(3) |
| 12975 | .nr(8) |
| 12976 | .kr(2) |
| 12977 | .sr(1) |
| 12978 | .m(m) |
| 12979 | .n(n) |
| 12980 | .k(k) |
| 12981 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12982 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 12983 | } |
| 12984 | } |
| 12985 | } |
| 12986 | } |
| 12987 | |
| 12988 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_div_16) { |
| 12989 | TEST_REQUIRES_ARM_NEON; |
| 12990 | for (size_t k = 32; k <= 160; k += 16) { |
| 12991 | GemmMicrokernelTester() |
| 12992 | .mr(3) |
| 12993 | .nr(8) |
| 12994 | .kr(2) |
| 12995 | .sr(1) |
| 12996 | .m(3) |
| 12997 | .n(8) |
| 12998 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 12999 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13000 | } |
| 13001 | } |
| 13002 | |
| 13003 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_div_16_strided_a) { |
| 13004 | TEST_REQUIRES_ARM_NEON; |
| 13005 | for (size_t k = 32; k <= 160; k += 16) { |
| 13006 | GemmMicrokernelTester() |
| 13007 | .mr(3) |
| 13008 | .nr(8) |
| 13009 | .kr(2) |
| 13010 | .sr(1) |
| 13011 | .m(3) |
| 13012 | .n(8) |
| 13013 | .k(k) |
| 13014 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13015 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13016 | } |
| 13017 | } |
| 13018 | |
| 13019 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_div_16_subtile) { |
| 13020 | TEST_REQUIRES_ARM_NEON; |
| 13021 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13022 | for (uint32_t n = 1; n <= 8; n++) { |
| 13023 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13024 | GemmMicrokernelTester() |
| 13025 | .mr(3) |
| 13026 | .nr(8) |
| 13027 | .kr(2) |
| 13028 | .sr(1) |
| 13029 | .m(m) |
| 13030 | .n(n) |
| 13031 | .k(k) |
| 13032 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13033 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13034 | } |
| 13035 | } |
| 13036 | } |
| 13037 | } |
| 13038 | |
| 13039 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8) { |
| 13040 | TEST_REQUIRES_ARM_NEON; |
| 13041 | for (uint32_t n = 9; n < 16; n++) { |
| 13042 | for (size_t k = 1; k <= 80; k += 17) { |
| 13043 | GemmMicrokernelTester() |
| 13044 | .mr(3) |
| 13045 | .nr(8) |
| 13046 | .kr(2) |
| 13047 | .sr(1) |
| 13048 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13049 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13050 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13051 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13052 | } |
| 13053 | } |
| 13054 | } |
| 13055 | |
| 13056 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) { |
| 13057 | TEST_REQUIRES_ARM_NEON; |
| 13058 | for (uint32_t n = 9; n < 16; n++) { |
| 13059 | for (size_t k = 1; k <= 80; k += 17) { |
| 13060 | GemmMicrokernelTester() |
| 13061 | .mr(3) |
| 13062 | .nr(8) |
| 13063 | .kr(2) |
| 13064 | .sr(1) |
| 13065 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13066 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13067 | .k(k) |
| 13068 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13069 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13070 | } |
| 13071 | } |
| 13072 | } |
| 13073 | |
| 13074 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_strided_a) { |
| 13075 | TEST_REQUIRES_ARM_NEON; |
| 13076 | for (uint32_t n = 9; n < 16; n++) { |
| 13077 | for (size_t k = 1; k <= 80; k += 17) { |
| 13078 | GemmMicrokernelTester() |
| 13079 | .mr(3) |
| 13080 | .nr(8) |
| 13081 | .kr(2) |
| 13082 | .sr(1) |
| 13083 | .m(3) |
| 13084 | .n(n) |
| 13085 | .k(k) |
| 13086 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13087 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13088 | } |
| 13089 | } |
| 13090 | } |
| 13091 | |
| 13092 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_subtile) { |
| 13093 | TEST_REQUIRES_ARM_NEON; |
| 13094 | for (uint32_t n = 9; n < 16; n++) { |
| 13095 | for (size_t k = 1; k <= 80; k += 17) { |
| 13096 | for (uint32_t m = 1; m <= 3; m++) { |
| 13097 | GemmMicrokernelTester() |
| 13098 | .mr(3) |
| 13099 | .nr(8) |
| 13100 | .kr(2) |
| 13101 | .sr(1) |
| 13102 | .m(m) |
| 13103 | .n(n) |
| 13104 | .k(k) |
| 13105 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13106 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13107 | } |
| 13108 | } |
| 13109 | } |
| 13110 | } |
| 13111 | |
| 13112 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8) { |
| 13113 | TEST_REQUIRES_ARM_NEON; |
| 13114 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13115 | for (size_t k = 1; k <= 80; k += 17) { |
| 13116 | GemmMicrokernelTester() |
| 13117 | .mr(3) |
| 13118 | .nr(8) |
| 13119 | .kr(2) |
| 13120 | .sr(1) |
| 13121 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13122 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13123 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13124 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13125 | } |
| 13126 | } |
| 13127 | } |
| 13128 | |
| 13129 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) { |
| 13130 | TEST_REQUIRES_ARM_NEON; |
| 13131 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13132 | for (size_t k = 1; k <= 80; k += 17) { |
| 13133 | GemmMicrokernelTester() |
| 13134 | .mr(3) |
| 13135 | .nr(8) |
| 13136 | .kr(2) |
| 13137 | .sr(1) |
| 13138 | .m(3) |
| 13139 | .n(n) |
| 13140 | .k(k) |
| 13141 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13142 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13143 | } |
| 13144 | } |
| 13145 | } |
| 13146 | |
| 13147 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_strided_a) { |
| 13148 | TEST_REQUIRES_ARM_NEON; |
| 13149 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13150 | for (size_t k = 1; k <= 80; k += 17) { |
| 13151 | GemmMicrokernelTester() |
| 13152 | .mr(3) |
| 13153 | .nr(8) |
| 13154 | .kr(2) |
| 13155 | .sr(1) |
| 13156 | .m(3) |
| 13157 | .n(n) |
| 13158 | .k(k) |
| 13159 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13160 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13161 | } |
| 13162 | } |
| 13163 | } |
| 13164 | |
| 13165 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_subtile) { |
| 13166 | TEST_REQUIRES_ARM_NEON; |
| 13167 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13168 | for (size_t k = 1; k <= 80; k += 17) { |
| 13169 | for (uint32_t m = 1; m <= 3; m++) { |
| 13170 | GemmMicrokernelTester() |
| 13171 | .mr(3) |
| 13172 | .nr(8) |
| 13173 | .kr(2) |
| 13174 | .sr(1) |
| 13175 | .m(m) |
| 13176 | .n(n) |
| 13177 | .k(k) |
| 13178 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13179 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13180 | } |
| 13181 | } |
| 13182 | } |
| 13183 | } |
| 13184 | |
| 13185 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cm_subtile) { |
| 13186 | TEST_REQUIRES_ARM_NEON; |
| 13187 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13188 | for (uint32_t n = 1; n <= 8; n++) { |
| 13189 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13190 | GemmMicrokernelTester() |
| 13191 | .mr(3) |
| 13192 | .nr(8) |
| 13193 | .kr(2) |
| 13194 | .sr(1) |
| 13195 | .m(m) |
| 13196 | .n(n) |
| 13197 | .k(k) |
| 13198 | .cm_stride(11) |
| 13199 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13200 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13201 | } |
| 13202 | } |
| 13203 | } |
| 13204 | } |
| 13205 | |
| 13206 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, qmin) { |
| 13207 | TEST_REQUIRES_ARM_NEON; |
| 13208 | GemmMicrokernelTester() |
| 13209 | .mr(3) |
| 13210 | .nr(8) |
| 13211 | .kr(2) |
| 13212 | .sr(1) |
| 13213 | .m(3) |
| 13214 | .n(8) |
| 13215 | .k(16) |
| 13216 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13217 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13218 | } |
| 13219 | |
| 13220 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, qmax) { |
| 13221 | TEST_REQUIRES_ARM_NEON; |
| 13222 | GemmMicrokernelTester() |
| 13223 | .mr(3) |
| 13224 | .nr(8) |
| 13225 | .kr(2) |
| 13226 | .sr(1) |
| 13227 | .m(3) |
| 13228 | .n(8) |
| 13229 | .k(16) |
| 13230 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13231 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13232 | } |
| 13233 | |
| 13234 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cm) { |
| 13235 | TEST_REQUIRES_ARM_NEON; |
| 13236 | GemmMicrokernelTester() |
| 13237 | .mr(3) |
| 13238 | .nr(8) |
| 13239 | .kr(2) |
| 13240 | .sr(1) |
| 13241 | .m(3) |
| 13242 | .n(8) |
| 13243 | .k(16) |
| 13244 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13245 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13246 | } |
| 13247 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 13248 | |
| 13249 | |
| 13250 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 13251 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_eq_16) { |
| 13252 | TEST_REQUIRES_ARM_NEON; |
| 13253 | GemmMicrokernelTester() |
| 13254 | .mr(4) |
| 13255 | .nr(8) |
| 13256 | .kr(2) |
| 13257 | .sr(1) |
| 13258 | .m(4) |
| 13259 | .n(8) |
| 13260 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13261 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13262 | } |
| 13263 | |
| 13264 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, strided_cn) { |
| 13265 | TEST_REQUIRES_ARM_NEON; |
| 13266 | GemmMicrokernelTester() |
| 13267 | .mr(4) |
| 13268 | .nr(8) |
| 13269 | .kr(2) |
| 13270 | .sr(1) |
| 13271 | .m(4) |
| 13272 | .n(8) |
| 13273 | .k(16) |
| 13274 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13275 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13276 | } |
| 13277 | |
| 13278 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_eq_16_strided_a) { |
| 13279 | TEST_REQUIRES_ARM_NEON; |
| 13280 | GemmMicrokernelTester() |
| 13281 | .mr(4) |
| 13282 | .nr(8) |
| 13283 | .kr(2) |
| 13284 | .sr(1) |
| 13285 | .m(4) |
| 13286 | .n(8) |
| 13287 | .k(16) |
| 13288 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13289 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13290 | } |
| 13291 | |
| 13292 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_eq_16_subtile) { |
| 13293 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13294 | for (uint32_t n = 1; n <= 8; n++) { |
| 13295 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13296 | GemmMicrokernelTester() |
| 13297 | .mr(4) |
| 13298 | .nr(8) |
| 13299 | .kr(2) |
| 13300 | .sr(1) |
| 13301 | .m(m) |
| 13302 | .n(n) |
| 13303 | .k(16) |
| 13304 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13305 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13306 | } |
| 13307 | } |
| 13308 | } |
| 13309 | |
| 13310 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) { |
| 13311 | TEST_REQUIRES_ARM_NEON; |
| 13312 | for (uint32_t m = 1; m <= 4; m++) { |
| 13313 | GemmMicrokernelTester() |
| 13314 | .mr(4) |
| 13315 | .nr(8) |
| 13316 | .kr(2) |
| 13317 | .sr(1) |
| 13318 | .m(m) |
| 13319 | .n(8) |
| 13320 | .k(16) |
| 13321 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13322 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13323 | } |
| 13324 | } |
| 13325 | |
| 13326 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) { |
| 13327 | TEST_REQUIRES_ARM_NEON; |
| 13328 | for (uint32_t n = 1; n <= 8; n++) { |
| 13329 | GemmMicrokernelTester() |
| 13330 | .mr(4) |
| 13331 | .nr(8) |
| 13332 | .kr(2) |
| 13333 | .sr(1) |
| 13334 | .m(4) |
| 13335 | .n(n) |
| 13336 | .k(16) |
| 13337 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13338 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13339 | } |
| 13340 | } |
| 13341 | |
| 13342 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_lt_16) { |
| 13343 | TEST_REQUIRES_ARM_NEON; |
| 13344 | for (size_t k = 1; k < 16; k++) { |
| 13345 | GemmMicrokernelTester() |
| 13346 | .mr(4) |
| 13347 | .nr(8) |
| 13348 | .kr(2) |
| 13349 | .sr(1) |
| 13350 | .m(4) |
| 13351 | .n(8) |
| 13352 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13353 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13354 | } |
| 13355 | } |
| 13356 | |
| 13357 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_lt_16_strided_a) { |
| 13358 | TEST_REQUIRES_ARM_NEON; |
| 13359 | for (size_t k = 1; k < 16; k++) { |
| 13360 | GemmMicrokernelTester() |
| 13361 | .mr(4) |
| 13362 | .nr(8) |
| 13363 | .kr(2) |
| 13364 | .sr(1) |
| 13365 | .m(4) |
| 13366 | .n(8) |
| 13367 | .k(k) |
| 13368 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13369 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13370 | } |
| 13371 | } |
| 13372 | |
| 13373 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_lt_16_subtile) { |
| 13374 | TEST_REQUIRES_ARM_NEON; |
| 13375 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13376 | for (uint32_t n = 1; n <= 8; n++) { |
| 13377 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13378 | GemmMicrokernelTester() |
| 13379 | .mr(4) |
| 13380 | .nr(8) |
| 13381 | .kr(2) |
| 13382 | .sr(1) |
| 13383 | .m(m) |
| 13384 | .n(n) |
| 13385 | .k(k) |
| 13386 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13387 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13388 | } |
| 13389 | } |
| 13390 | } |
| 13391 | } |
| 13392 | |
| 13393 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_gt_16) { |
| 13394 | TEST_REQUIRES_ARM_NEON; |
| 13395 | for (size_t k = 17; k < 32; k++) { |
| 13396 | GemmMicrokernelTester() |
| 13397 | .mr(4) |
| 13398 | .nr(8) |
| 13399 | .kr(2) |
| 13400 | .sr(1) |
| 13401 | .m(4) |
| 13402 | .n(8) |
| 13403 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13404 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13405 | } |
| 13406 | } |
| 13407 | |
| 13408 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_gt_16_strided_a) { |
| 13409 | TEST_REQUIRES_ARM_NEON; |
| 13410 | for (size_t k = 17; k < 32; k++) { |
| 13411 | GemmMicrokernelTester() |
| 13412 | .mr(4) |
| 13413 | .nr(8) |
| 13414 | .kr(2) |
| 13415 | .sr(1) |
| 13416 | .m(4) |
| 13417 | .n(8) |
| 13418 | .k(k) |
| 13419 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13420 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13421 | } |
| 13422 | } |
| 13423 | |
| 13424 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_gt_16_subtile) { |
| 13425 | TEST_REQUIRES_ARM_NEON; |
| 13426 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13427 | for (uint32_t n = 1; n <= 8; n++) { |
| 13428 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13429 | GemmMicrokernelTester() |
| 13430 | .mr(4) |
| 13431 | .nr(8) |
| 13432 | .kr(2) |
| 13433 | .sr(1) |
| 13434 | .m(m) |
| 13435 | .n(n) |
| 13436 | .k(k) |
| 13437 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13438 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13439 | } |
| 13440 | } |
| 13441 | } |
| 13442 | } |
| 13443 | |
| 13444 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_div_16) { |
| 13445 | TEST_REQUIRES_ARM_NEON; |
| 13446 | for (size_t k = 32; k <= 160; k += 16) { |
| 13447 | GemmMicrokernelTester() |
| 13448 | .mr(4) |
| 13449 | .nr(8) |
| 13450 | .kr(2) |
| 13451 | .sr(1) |
| 13452 | .m(4) |
| 13453 | .n(8) |
| 13454 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13455 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13456 | } |
| 13457 | } |
| 13458 | |
| 13459 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_div_16_strided_a) { |
| 13460 | TEST_REQUIRES_ARM_NEON; |
| 13461 | for (size_t k = 32; k <= 160; k += 16) { |
| 13462 | GemmMicrokernelTester() |
| 13463 | .mr(4) |
| 13464 | .nr(8) |
| 13465 | .kr(2) |
| 13466 | .sr(1) |
| 13467 | .m(4) |
| 13468 | .n(8) |
| 13469 | .k(k) |
| 13470 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13471 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13472 | } |
| 13473 | } |
| 13474 | |
| 13475 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, k_div_16_subtile) { |
| 13476 | TEST_REQUIRES_ARM_NEON; |
| 13477 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13478 | for (uint32_t n = 1; n <= 8; n++) { |
| 13479 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13480 | GemmMicrokernelTester() |
| 13481 | .mr(4) |
| 13482 | .nr(8) |
| 13483 | .kr(2) |
| 13484 | .sr(1) |
| 13485 | .m(m) |
| 13486 | .n(n) |
| 13487 | .k(k) |
| 13488 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13489 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13490 | } |
| 13491 | } |
| 13492 | } |
| 13493 | } |
| 13494 | |
| 13495 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_gt_8) { |
| 13496 | TEST_REQUIRES_ARM_NEON; |
| 13497 | for (uint32_t n = 9; n < 16; n++) { |
| 13498 | for (size_t k = 1; k <= 80; k += 17) { |
| 13499 | GemmMicrokernelTester() |
| 13500 | .mr(4) |
| 13501 | .nr(8) |
| 13502 | .kr(2) |
| 13503 | .sr(1) |
| 13504 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13505 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13506 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13507 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13508 | } |
| 13509 | } |
| 13510 | } |
| 13511 | |
| 13512 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) { |
| 13513 | TEST_REQUIRES_ARM_NEON; |
| 13514 | for (uint32_t n = 9; n < 16; n++) { |
| 13515 | for (size_t k = 1; k <= 80; k += 17) { |
| 13516 | GemmMicrokernelTester() |
| 13517 | .mr(4) |
| 13518 | .nr(8) |
| 13519 | .kr(2) |
| 13520 | .sr(1) |
| 13521 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13522 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13523 | .k(k) |
| 13524 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13525 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13526 | } |
| 13527 | } |
| 13528 | } |
| 13529 | |
| 13530 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_gt_8_strided_a) { |
| 13531 | TEST_REQUIRES_ARM_NEON; |
| 13532 | for (uint32_t n = 9; n < 16; n++) { |
| 13533 | for (size_t k = 1; k <= 80; k += 17) { |
| 13534 | GemmMicrokernelTester() |
| 13535 | .mr(4) |
| 13536 | .nr(8) |
| 13537 | .kr(2) |
| 13538 | .sr(1) |
| 13539 | .m(4) |
| 13540 | .n(n) |
| 13541 | .k(k) |
| 13542 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13543 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13544 | } |
| 13545 | } |
| 13546 | } |
| 13547 | |
| 13548 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_gt_8_subtile) { |
| 13549 | TEST_REQUIRES_ARM_NEON; |
| 13550 | for (uint32_t n = 9; n < 16; n++) { |
| 13551 | for (size_t k = 1; k <= 80; k += 17) { |
| 13552 | for (uint32_t m = 1; m <= 4; m++) { |
| 13553 | GemmMicrokernelTester() |
| 13554 | .mr(4) |
| 13555 | .nr(8) |
| 13556 | .kr(2) |
| 13557 | .sr(1) |
| 13558 | .m(m) |
| 13559 | .n(n) |
| 13560 | .k(k) |
| 13561 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13562 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13563 | } |
| 13564 | } |
| 13565 | } |
| 13566 | } |
| 13567 | |
| 13568 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_div_8) { |
| 13569 | TEST_REQUIRES_ARM_NEON; |
| 13570 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13571 | for (size_t k = 1; k <= 80; k += 17) { |
| 13572 | GemmMicrokernelTester() |
| 13573 | .mr(4) |
| 13574 | .nr(8) |
| 13575 | .kr(2) |
| 13576 | .sr(1) |
| 13577 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13578 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13579 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13580 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13581 | } |
| 13582 | } |
| 13583 | } |
| 13584 | |
| 13585 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) { |
| 13586 | TEST_REQUIRES_ARM_NEON; |
| 13587 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13588 | for (size_t k = 1; k <= 80; k += 17) { |
| 13589 | GemmMicrokernelTester() |
| 13590 | .mr(4) |
| 13591 | .nr(8) |
| 13592 | .kr(2) |
| 13593 | .sr(1) |
| 13594 | .m(4) |
| 13595 | .n(n) |
| 13596 | .k(k) |
| 13597 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13598 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13599 | } |
| 13600 | } |
| 13601 | } |
| 13602 | |
| 13603 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_div_8_strided_a) { |
| 13604 | TEST_REQUIRES_ARM_NEON; |
| 13605 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13606 | for (size_t k = 1; k <= 80; k += 17) { |
| 13607 | GemmMicrokernelTester() |
| 13608 | .mr(4) |
| 13609 | .nr(8) |
| 13610 | .kr(2) |
| 13611 | .sr(1) |
| 13612 | .m(4) |
| 13613 | .n(n) |
| 13614 | .k(k) |
| 13615 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13616 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13617 | } |
| 13618 | } |
| 13619 | } |
| 13620 | |
| 13621 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, n_div_8_subtile) { |
| 13622 | TEST_REQUIRES_ARM_NEON; |
| 13623 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 13624 | for (size_t k = 1; k <= 80; k += 17) { |
| 13625 | for (uint32_t m = 1; m <= 4; m++) { |
| 13626 | GemmMicrokernelTester() |
| 13627 | .mr(4) |
| 13628 | .nr(8) |
| 13629 | .kr(2) |
| 13630 | .sr(1) |
| 13631 | .m(m) |
| 13632 | .n(n) |
| 13633 | .k(k) |
| 13634 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13635 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13636 | } |
| 13637 | } |
| 13638 | } |
| 13639 | } |
| 13640 | |
| 13641 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, strided_cm_subtile) { |
| 13642 | TEST_REQUIRES_ARM_NEON; |
| 13643 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13644 | for (uint32_t n = 1; n <= 8; n++) { |
| 13645 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13646 | GemmMicrokernelTester() |
| 13647 | .mr(4) |
| 13648 | .nr(8) |
| 13649 | .kr(2) |
| 13650 | .sr(1) |
| 13651 | .m(m) |
| 13652 | .n(n) |
| 13653 | .k(k) |
| 13654 | .cm_stride(11) |
| 13655 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13656 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13657 | } |
| 13658 | } |
| 13659 | } |
| 13660 | } |
| 13661 | |
| 13662 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, qmin) { |
| 13663 | TEST_REQUIRES_ARM_NEON; |
| 13664 | GemmMicrokernelTester() |
| 13665 | .mr(4) |
| 13666 | .nr(8) |
| 13667 | .kr(2) |
| 13668 | .sr(1) |
| 13669 | .m(4) |
| 13670 | .n(8) |
| 13671 | .k(16) |
| 13672 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13673 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13674 | } |
| 13675 | |
| 13676 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, qmax) { |
| 13677 | TEST_REQUIRES_ARM_NEON; |
| 13678 | GemmMicrokernelTester() |
| 13679 | .mr(4) |
| 13680 | .nr(8) |
| 13681 | .kr(2) |
| 13682 | .sr(1) |
| 13683 | .m(4) |
| 13684 | .n(8) |
| 13685 | .k(16) |
| 13686 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13687 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13688 | } |
| 13689 | |
| 13690 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_DUP, strided_cm) { |
| 13691 | TEST_REQUIRES_ARM_NEON; |
| 13692 | GemmMicrokernelTester() |
| 13693 | .mr(4) |
| 13694 | .nr(8) |
| 13695 | .kr(2) |
| 13696 | .sr(1) |
| 13697 | .m(4) |
| 13698 | .n(8) |
| 13699 | .k(16) |
| 13700 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13701 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13702 | } |
| 13703 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 13704 | |
| 13705 | |
| 13706 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 13707 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16) { |
| 13708 | TEST_REQUIRES_ARM_NEON; |
| 13709 | GemmMicrokernelTester() |
| 13710 | .mr(1) |
| 13711 | .nr(16) |
| 13712 | .kr(2) |
| 13713 | .sr(1) |
| 13714 | .m(1) |
| 13715 | .n(16) |
| 13716 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13717 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13718 | } |
| 13719 | |
| 13720 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, strided_cn) { |
| 13721 | TEST_REQUIRES_ARM_NEON; |
| 13722 | GemmMicrokernelTester() |
| 13723 | .mr(1) |
| 13724 | .nr(16) |
| 13725 | .kr(2) |
| 13726 | .sr(1) |
| 13727 | .m(1) |
| 13728 | .n(16) |
| 13729 | .k(16) |
| 13730 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13731 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13732 | } |
| 13733 | |
| 13734 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16_strided_a) { |
| 13735 | TEST_REQUIRES_ARM_NEON; |
| 13736 | GemmMicrokernelTester() |
| 13737 | .mr(1) |
| 13738 | .nr(16) |
| 13739 | .kr(2) |
| 13740 | .sr(1) |
| 13741 | .m(1) |
| 13742 | .n(16) |
| 13743 | .k(16) |
| 13744 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13745 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13746 | } |
| 13747 | |
| 13748 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16_subtile) { |
| 13749 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13750 | for (uint32_t n = 1; n <= 16; n++) { |
| 13751 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13752 | GemmMicrokernelTester() |
| 13753 | .mr(1) |
| 13754 | .nr(16) |
| 13755 | .kr(2) |
| 13756 | .sr(1) |
| 13757 | .m(m) |
| 13758 | .n(n) |
| 13759 | .k(16) |
| 13760 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13761 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13762 | } |
| 13763 | } |
| 13764 | } |
| 13765 | |
| 13766 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16_subtile_m) { |
| 13767 | TEST_REQUIRES_ARM_NEON; |
| 13768 | for (uint32_t m = 1; m <= 1; m++) { |
| 13769 | GemmMicrokernelTester() |
| 13770 | .mr(1) |
| 13771 | .nr(16) |
| 13772 | .kr(2) |
| 13773 | .sr(1) |
| 13774 | .m(m) |
| 13775 | .n(16) |
| 13776 | .k(16) |
| 13777 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13778 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13779 | } |
| 13780 | } |
| 13781 | |
| 13782 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16_subtile_n) { |
| 13783 | TEST_REQUIRES_ARM_NEON; |
| 13784 | for (uint32_t n = 1; n <= 16; n++) { |
| 13785 | GemmMicrokernelTester() |
| 13786 | .mr(1) |
| 13787 | .nr(16) |
| 13788 | .kr(2) |
| 13789 | .sr(1) |
| 13790 | .m(1) |
| 13791 | .n(n) |
| 13792 | .k(16) |
| 13793 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13794 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13795 | } |
| 13796 | } |
| 13797 | |
| 13798 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_lt_16) { |
| 13799 | TEST_REQUIRES_ARM_NEON; |
| 13800 | for (size_t k = 1; k < 16; k++) { |
| 13801 | GemmMicrokernelTester() |
| 13802 | .mr(1) |
| 13803 | .nr(16) |
| 13804 | .kr(2) |
| 13805 | .sr(1) |
| 13806 | .m(1) |
| 13807 | .n(16) |
| 13808 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13809 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13810 | } |
| 13811 | } |
| 13812 | |
| 13813 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_lt_16_strided_a) { |
| 13814 | TEST_REQUIRES_ARM_NEON; |
| 13815 | for (size_t k = 1; k < 16; k++) { |
| 13816 | GemmMicrokernelTester() |
| 13817 | .mr(1) |
| 13818 | .nr(16) |
| 13819 | .kr(2) |
| 13820 | .sr(1) |
| 13821 | .m(1) |
| 13822 | .n(16) |
| 13823 | .k(k) |
| 13824 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13825 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13826 | } |
| 13827 | } |
| 13828 | |
| 13829 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_lt_16_subtile) { |
| 13830 | TEST_REQUIRES_ARM_NEON; |
| 13831 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13832 | for (uint32_t n = 1; n <= 16; n++) { |
| 13833 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13834 | GemmMicrokernelTester() |
| 13835 | .mr(1) |
| 13836 | .nr(16) |
| 13837 | .kr(2) |
| 13838 | .sr(1) |
| 13839 | .m(m) |
| 13840 | .n(n) |
| 13841 | .k(k) |
| 13842 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13843 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13844 | } |
| 13845 | } |
| 13846 | } |
| 13847 | } |
| 13848 | |
| 13849 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_gt_16) { |
| 13850 | TEST_REQUIRES_ARM_NEON; |
| 13851 | for (size_t k = 17; k < 32; k++) { |
| 13852 | GemmMicrokernelTester() |
| 13853 | .mr(1) |
| 13854 | .nr(16) |
| 13855 | .kr(2) |
| 13856 | .sr(1) |
| 13857 | .m(1) |
| 13858 | .n(16) |
| 13859 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13860 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13861 | } |
| 13862 | } |
| 13863 | |
| 13864 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_gt_16_strided_a) { |
| 13865 | TEST_REQUIRES_ARM_NEON; |
| 13866 | for (size_t k = 17; k < 32; k++) { |
| 13867 | GemmMicrokernelTester() |
| 13868 | .mr(1) |
| 13869 | .nr(16) |
| 13870 | .kr(2) |
| 13871 | .sr(1) |
| 13872 | .m(1) |
| 13873 | .n(16) |
| 13874 | .k(k) |
| 13875 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13876 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13877 | } |
| 13878 | } |
| 13879 | |
| 13880 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_gt_16_subtile) { |
| 13881 | TEST_REQUIRES_ARM_NEON; |
| 13882 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13883 | for (uint32_t n = 1; n <= 16; n++) { |
| 13884 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13885 | GemmMicrokernelTester() |
| 13886 | .mr(1) |
| 13887 | .nr(16) |
| 13888 | .kr(2) |
| 13889 | .sr(1) |
| 13890 | .m(m) |
| 13891 | .n(n) |
| 13892 | .k(k) |
| 13893 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13894 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13895 | } |
| 13896 | } |
| 13897 | } |
| 13898 | } |
| 13899 | |
| 13900 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_div_16) { |
| 13901 | TEST_REQUIRES_ARM_NEON; |
| 13902 | for (size_t k = 32; k <= 160; k += 16) { |
| 13903 | GemmMicrokernelTester() |
| 13904 | .mr(1) |
| 13905 | .nr(16) |
| 13906 | .kr(2) |
| 13907 | .sr(1) |
| 13908 | .m(1) |
| 13909 | .n(16) |
| 13910 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13911 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13912 | } |
| 13913 | } |
| 13914 | |
| 13915 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_div_16_strided_a) { |
| 13916 | TEST_REQUIRES_ARM_NEON; |
| 13917 | for (size_t k = 32; k <= 160; k += 16) { |
| 13918 | GemmMicrokernelTester() |
| 13919 | .mr(1) |
| 13920 | .nr(16) |
| 13921 | .kr(2) |
| 13922 | .sr(1) |
| 13923 | .m(1) |
| 13924 | .n(16) |
| 13925 | .k(k) |
| 13926 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13927 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13928 | } |
| 13929 | } |
| 13930 | |
| 13931 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_div_16_subtile) { |
| 13932 | TEST_REQUIRES_ARM_NEON; |
| 13933 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 13934 | for (uint32_t n = 1; n <= 16; n++) { |
| 13935 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13936 | GemmMicrokernelTester() |
| 13937 | .mr(1) |
| 13938 | .nr(16) |
| 13939 | .kr(2) |
| 13940 | .sr(1) |
| 13941 | .m(m) |
| 13942 | .n(n) |
| 13943 | .k(k) |
| 13944 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13945 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13946 | } |
| 13947 | } |
| 13948 | } |
| 13949 | } |
| 13950 | |
| 13951 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16) { |
| 13952 | TEST_REQUIRES_ARM_NEON; |
| 13953 | for (uint32_t n = 17; n < 32; n++) { |
| 13954 | for (size_t k = 1; k <= 80; k += 17) { |
| 13955 | GemmMicrokernelTester() |
| 13956 | .mr(1) |
| 13957 | .nr(16) |
| 13958 | .kr(2) |
| 13959 | .sr(1) |
| 13960 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13961 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13962 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13963 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13964 | } |
| 13965 | } |
| 13966 | } |
| 13967 | |
| 13968 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16_strided_cn) { |
| 13969 | TEST_REQUIRES_ARM_NEON; |
| 13970 | for (uint32_t n = 17; n < 32; n++) { |
| 13971 | for (size_t k = 1; k <= 80; k += 17) { |
| 13972 | GemmMicrokernelTester() |
| 13973 | .mr(1) |
| 13974 | .nr(16) |
| 13975 | .kr(2) |
| 13976 | .sr(1) |
| 13977 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 13978 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13979 | .k(k) |
| 13980 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13981 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 13982 | } |
| 13983 | } |
| 13984 | } |
| 13985 | |
| 13986 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16_strided_a) { |
| 13987 | TEST_REQUIRES_ARM_NEON; |
| 13988 | for (uint32_t n = 17; n < 32; n++) { |
| 13989 | for (size_t k = 1; k <= 80; k += 17) { |
| 13990 | GemmMicrokernelTester() |
| 13991 | .mr(1) |
| 13992 | .nr(16) |
| 13993 | .kr(2) |
| 13994 | .sr(1) |
| 13995 | .m(1) |
| 13996 | .n(n) |
| 13997 | .k(k) |
| 13998 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 13999 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14000 | } |
| 14001 | } |
| 14002 | } |
| 14003 | |
| 14004 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16_subtile) { |
| 14005 | TEST_REQUIRES_ARM_NEON; |
| 14006 | for (uint32_t n = 17; n < 32; n++) { |
| 14007 | for (size_t k = 1; k <= 80; k += 17) { |
| 14008 | for (uint32_t m = 1; m <= 1; m++) { |
| 14009 | GemmMicrokernelTester() |
| 14010 | .mr(1) |
| 14011 | .nr(16) |
| 14012 | .kr(2) |
| 14013 | .sr(1) |
| 14014 | .m(m) |
| 14015 | .n(n) |
| 14016 | .k(k) |
| 14017 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14018 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14019 | } |
| 14020 | } |
| 14021 | } |
| 14022 | } |
| 14023 | |
| 14024 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16) { |
| 14025 | TEST_REQUIRES_ARM_NEON; |
| 14026 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 14027 | for (size_t k = 1; k <= 80; k += 17) { |
| 14028 | GemmMicrokernelTester() |
| 14029 | .mr(1) |
| 14030 | .nr(16) |
| 14031 | .kr(2) |
| 14032 | .sr(1) |
| 14033 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 14034 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14035 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14036 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14037 | } |
| 14038 | } |
| 14039 | } |
| 14040 | |
| 14041 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16_strided_cn) { |
| 14042 | TEST_REQUIRES_ARM_NEON; |
| 14043 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 14044 | for (size_t k = 1; k <= 80; k += 17) { |
| 14045 | GemmMicrokernelTester() |
| 14046 | .mr(1) |
| 14047 | .nr(16) |
| 14048 | .kr(2) |
| 14049 | .sr(1) |
| 14050 | .m(1) |
| 14051 | .n(n) |
| 14052 | .k(k) |
| 14053 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14054 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14055 | } |
| 14056 | } |
| 14057 | } |
| 14058 | |
| 14059 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16_strided_a) { |
| 14060 | TEST_REQUIRES_ARM_NEON; |
| 14061 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 14062 | for (size_t k = 1; k <= 80; k += 17) { |
| 14063 | GemmMicrokernelTester() |
| 14064 | .mr(1) |
| 14065 | .nr(16) |
| 14066 | .kr(2) |
| 14067 | .sr(1) |
| 14068 | .m(1) |
| 14069 | .n(n) |
| 14070 | .k(k) |
| 14071 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14072 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14073 | } |
| 14074 | } |
| 14075 | } |
| 14076 | |
| 14077 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16_subtile) { |
| 14078 | TEST_REQUIRES_ARM_NEON; |
| 14079 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 14080 | for (size_t k = 1; k <= 80; k += 17) { |
| 14081 | for (uint32_t m = 1; m <= 1; m++) { |
| 14082 | GemmMicrokernelTester() |
| 14083 | .mr(1) |
| 14084 | .nr(16) |
| 14085 | .kr(2) |
| 14086 | .sr(1) |
| 14087 | .m(m) |
| 14088 | .n(n) |
| 14089 | .k(k) |
| 14090 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14091 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14092 | } |
| 14093 | } |
| 14094 | } |
| 14095 | } |
| 14096 | |
| 14097 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, strided_cm_subtile) { |
| 14098 | TEST_REQUIRES_ARM_NEON; |
| 14099 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14100 | for (uint32_t n = 1; n <= 16; n++) { |
| 14101 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14102 | GemmMicrokernelTester() |
| 14103 | .mr(1) |
| 14104 | .nr(16) |
| 14105 | .kr(2) |
| 14106 | .sr(1) |
| 14107 | .m(m) |
| 14108 | .n(n) |
| 14109 | .k(k) |
| 14110 | .cm_stride(19) |
| 14111 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14112 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14113 | } |
| 14114 | } |
| 14115 | } |
| 14116 | } |
| 14117 | |
| 14118 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, qmin) { |
| 14119 | TEST_REQUIRES_ARM_NEON; |
| 14120 | GemmMicrokernelTester() |
| 14121 | .mr(1) |
| 14122 | .nr(16) |
| 14123 | .kr(2) |
| 14124 | .sr(1) |
| 14125 | .m(1) |
| 14126 | .n(16) |
| 14127 | .k(16) |
| 14128 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14129 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14130 | } |
| 14131 | |
| 14132 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, qmax) { |
| 14133 | TEST_REQUIRES_ARM_NEON; |
| 14134 | GemmMicrokernelTester() |
| 14135 | .mr(1) |
| 14136 | .nr(16) |
| 14137 | .kr(2) |
| 14138 | .sr(1) |
| 14139 | .m(1) |
| 14140 | .n(16) |
| 14141 | .k(16) |
| 14142 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14143 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14144 | } |
| 14145 | |
| 14146 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, strided_cm) { |
| 14147 | TEST_REQUIRES_ARM_NEON; |
| 14148 | GemmMicrokernelTester() |
| 14149 | .mr(1) |
| 14150 | .nr(16) |
| 14151 | .kr(2) |
| 14152 | .sr(1) |
| 14153 | .m(1) |
| 14154 | .n(16) |
| 14155 | .k(16) |
| 14156 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14157 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14158 | } |
| 14159 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 14160 | |
| 14161 | |
| 14162 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 14163 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_eq_16) { |
| 14164 | TEST_REQUIRES_ARM_NEON; |
| 14165 | GemmMicrokernelTester() |
| 14166 | .mr(2) |
| 14167 | .nr(8) |
| 14168 | .kr(8) |
| 14169 | .sr(1) |
| 14170 | .m(2) |
| 14171 | .n(8) |
| 14172 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14173 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14174 | } |
| 14175 | |
| 14176 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, strided_cn) { |
| 14177 | TEST_REQUIRES_ARM_NEON; |
| 14178 | GemmMicrokernelTester() |
| 14179 | .mr(2) |
| 14180 | .nr(8) |
| 14181 | .kr(8) |
| 14182 | .sr(1) |
| 14183 | .m(2) |
| 14184 | .n(8) |
| 14185 | .k(16) |
| 14186 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14187 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14188 | } |
| 14189 | |
| 14190 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_eq_16_strided_a) { |
| 14191 | TEST_REQUIRES_ARM_NEON; |
| 14192 | GemmMicrokernelTester() |
| 14193 | .mr(2) |
| 14194 | .nr(8) |
| 14195 | .kr(8) |
| 14196 | .sr(1) |
| 14197 | .m(2) |
| 14198 | .n(8) |
| 14199 | .k(16) |
| 14200 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14201 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14202 | } |
| 14203 | |
| 14204 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_eq_16_subtile) { |
| 14205 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14206 | for (uint32_t n = 1; n <= 8; n++) { |
| 14207 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14208 | GemmMicrokernelTester() |
| 14209 | .mr(2) |
| 14210 | .nr(8) |
| 14211 | .kr(8) |
| 14212 | .sr(1) |
| 14213 | .m(m) |
| 14214 | .n(n) |
| 14215 | .k(16) |
| 14216 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14217 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14218 | } |
| 14219 | } |
| 14220 | } |
| 14221 | |
| 14222 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_eq_16_subtile_m) { |
| 14223 | TEST_REQUIRES_ARM_NEON; |
| 14224 | for (uint32_t m = 1; m <= 2; m++) { |
| 14225 | GemmMicrokernelTester() |
| 14226 | .mr(2) |
| 14227 | .nr(8) |
| 14228 | .kr(8) |
| 14229 | .sr(1) |
| 14230 | .m(m) |
| 14231 | .n(8) |
| 14232 | .k(16) |
| 14233 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14234 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14235 | } |
| 14236 | } |
| 14237 | |
| 14238 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_eq_16_subtile_n) { |
| 14239 | TEST_REQUIRES_ARM_NEON; |
| 14240 | for (uint32_t n = 1; n <= 8; n++) { |
| 14241 | GemmMicrokernelTester() |
| 14242 | .mr(2) |
| 14243 | .nr(8) |
| 14244 | .kr(8) |
| 14245 | .sr(1) |
| 14246 | .m(2) |
| 14247 | .n(n) |
| 14248 | .k(16) |
| 14249 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14250 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14251 | } |
| 14252 | } |
| 14253 | |
| 14254 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_lt_16) { |
| 14255 | TEST_REQUIRES_ARM_NEON; |
| 14256 | for (size_t k = 1; k < 16; k++) { |
| 14257 | GemmMicrokernelTester() |
| 14258 | .mr(2) |
| 14259 | .nr(8) |
| 14260 | .kr(8) |
| 14261 | .sr(1) |
| 14262 | .m(2) |
| 14263 | .n(8) |
| 14264 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14265 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14266 | } |
| 14267 | } |
| 14268 | |
| 14269 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_lt_16_strided_a) { |
| 14270 | TEST_REQUIRES_ARM_NEON; |
| 14271 | for (size_t k = 1; k < 16; k++) { |
| 14272 | GemmMicrokernelTester() |
| 14273 | .mr(2) |
| 14274 | .nr(8) |
| 14275 | .kr(8) |
| 14276 | .sr(1) |
| 14277 | .m(2) |
| 14278 | .n(8) |
| 14279 | .k(k) |
| 14280 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14281 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14282 | } |
| 14283 | } |
| 14284 | |
| 14285 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_lt_16_subtile) { |
| 14286 | TEST_REQUIRES_ARM_NEON; |
| 14287 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14288 | for (uint32_t n = 1; n <= 8; n++) { |
| 14289 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14290 | GemmMicrokernelTester() |
| 14291 | .mr(2) |
| 14292 | .nr(8) |
| 14293 | .kr(8) |
| 14294 | .sr(1) |
| 14295 | .m(m) |
| 14296 | .n(n) |
| 14297 | .k(k) |
| 14298 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14299 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14300 | } |
| 14301 | } |
| 14302 | } |
| 14303 | } |
| 14304 | |
| 14305 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_gt_16) { |
| 14306 | TEST_REQUIRES_ARM_NEON; |
| 14307 | for (size_t k = 17; k < 32; k++) { |
| 14308 | GemmMicrokernelTester() |
| 14309 | .mr(2) |
| 14310 | .nr(8) |
| 14311 | .kr(8) |
| 14312 | .sr(1) |
| 14313 | .m(2) |
| 14314 | .n(8) |
| 14315 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14316 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14317 | } |
| 14318 | } |
| 14319 | |
| 14320 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_gt_16_strided_a) { |
| 14321 | TEST_REQUIRES_ARM_NEON; |
| 14322 | for (size_t k = 17; k < 32; k++) { |
| 14323 | GemmMicrokernelTester() |
| 14324 | .mr(2) |
| 14325 | .nr(8) |
| 14326 | .kr(8) |
| 14327 | .sr(1) |
| 14328 | .m(2) |
| 14329 | .n(8) |
| 14330 | .k(k) |
| 14331 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14332 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14333 | } |
| 14334 | } |
| 14335 | |
| 14336 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_gt_16_subtile) { |
| 14337 | TEST_REQUIRES_ARM_NEON; |
| 14338 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14339 | for (uint32_t n = 1; n <= 8; n++) { |
| 14340 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14341 | GemmMicrokernelTester() |
| 14342 | .mr(2) |
| 14343 | .nr(8) |
| 14344 | .kr(8) |
| 14345 | .sr(1) |
| 14346 | .m(m) |
| 14347 | .n(n) |
| 14348 | .k(k) |
| 14349 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14350 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14351 | } |
| 14352 | } |
| 14353 | } |
| 14354 | } |
| 14355 | |
| 14356 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_div_16) { |
| 14357 | TEST_REQUIRES_ARM_NEON; |
| 14358 | for (size_t k = 32; k <= 160; k += 16) { |
| 14359 | GemmMicrokernelTester() |
| 14360 | .mr(2) |
| 14361 | .nr(8) |
| 14362 | .kr(8) |
| 14363 | .sr(1) |
| 14364 | .m(2) |
| 14365 | .n(8) |
| 14366 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14367 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14368 | } |
| 14369 | } |
| 14370 | |
| 14371 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_div_16_strided_a) { |
| 14372 | TEST_REQUIRES_ARM_NEON; |
| 14373 | for (size_t k = 32; k <= 160; k += 16) { |
| 14374 | GemmMicrokernelTester() |
| 14375 | .mr(2) |
| 14376 | .nr(8) |
| 14377 | .kr(8) |
| 14378 | .sr(1) |
| 14379 | .m(2) |
| 14380 | .n(8) |
| 14381 | .k(k) |
| 14382 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14383 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14384 | } |
| 14385 | } |
| 14386 | |
| 14387 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, k_div_16_subtile) { |
| 14388 | TEST_REQUIRES_ARM_NEON; |
| 14389 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14390 | for (uint32_t n = 1; n <= 8; n++) { |
| 14391 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14392 | GemmMicrokernelTester() |
| 14393 | .mr(2) |
| 14394 | .nr(8) |
| 14395 | .kr(8) |
| 14396 | .sr(1) |
| 14397 | .m(m) |
| 14398 | .n(n) |
| 14399 | .k(k) |
| 14400 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14401 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14402 | } |
| 14403 | } |
| 14404 | } |
| 14405 | } |
| 14406 | |
| 14407 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_gt_8) { |
| 14408 | TEST_REQUIRES_ARM_NEON; |
| 14409 | for (uint32_t n = 9; n < 16; n++) { |
| 14410 | for (size_t k = 1; k <= 80; k += 17) { |
| 14411 | GemmMicrokernelTester() |
| 14412 | .mr(2) |
| 14413 | .nr(8) |
| 14414 | .kr(8) |
| 14415 | .sr(1) |
| 14416 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 14417 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14418 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14419 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14420 | } |
| 14421 | } |
| 14422 | } |
| 14423 | |
| 14424 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_gt_8_strided_cn) { |
| 14425 | TEST_REQUIRES_ARM_NEON; |
| 14426 | for (uint32_t n = 9; n < 16; n++) { |
| 14427 | for (size_t k = 1; k <= 80; k += 17) { |
| 14428 | GemmMicrokernelTester() |
| 14429 | .mr(2) |
| 14430 | .nr(8) |
| 14431 | .kr(8) |
| 14432 | .sr(1) |
| 14433 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 14434 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14435 | .k(k) |
| 14436 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14437 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14438 | } |
| 14439 | } |
| 14440 | } |
| 14441 | |
| 14442 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_gt_8_strided_a) { |
| 14443 | TEST_REQUIRES_ARM_NEON; |
| 14444 | for (uint32_t n = 9; n < 16; n++) { |
| 14445 | for (size_t k = 1; k <= 80; k += 17) { |
| 14446 | GemmMicrokernelTester() |
| 14447 | .mr(2) |
| 14448 | .nr(8) |
| 14449 | .kr(8) |
| 14450 | .sr(1) |
| 14451 | .m(2) |
| 14452 | .n(n) |
| 14453 | .k(k) |
| 14454 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14455 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14456 | } |
| 14457 | } |
| 14458 | } |
| 14459 | |
| 14460 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_gt_8_subtile) { |
| 14461 | TEST_REQUIRES_ARM_NEON; |
| 14462 | for (uint32_t n = 9; n < 16; n++) { |
| 14463 | for (size_t k = 1; k <= 80; k += 17) { |
| 14464 | for (uint32_t m = 1; m <= 2; m++) { |
| 14465 | GemmMicrokernelTester() |
| 14466 | .mr(2) |
| 14467 | .nr(8) |
| 14468 | .kr(8) |
| 14469 | .sr(1) |
| 14470 | .m(m) |
| 14471 | .n(n) |
| 14472 | .k(k) |
| 14473 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14474 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14475 | } |
| 14476 | } |
| 14477 | } |
| 14478 | } |
| 14479 | |
| 14480 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_div_8) { |
| 14481 | TEST_REQUIRES_ARM_NEON; |
| 14482 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14483 | for (size_t k = 1; k <= 80; k += 17) { |
| 14484 | GemmMicrokernelTester() |
| 14485 | .mr(2) |
| 14486 | .nr(8) |
| 14487 | .kr(8) |
| 14488 | .sr(1) |
| 14489 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 14490 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14491 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14492 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14493 | } |
| 14494 | } |
| 14495 | } |
| 14496 | |
| 14497 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_div_8_strided_cn) { |
| 14498 | TEST_REQUIRES_ARM_NEON; |
| 14499 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14500 | for (size_t k = 1; k <= 80; k += 17) { |
| 14501 | GemmMicrokernelTester() |
| 14502 | .mr(2) |
| 14503 | .nr(8) |
| 14504 | .kr(8) |
| 14505 | .sr(1) |
| 14506 | .m(2) |
| 14507 | .n(n) |
| 14508 | .k(k) |
| 14509 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14510 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14511 | } |
| 14512 | } |
| 14513 | } |
| 14514 | |
| 14515 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_div_8_strided_a) { |
| 14516 | TEST_REQUIRES_ARM_NEON; |
| 14517 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14518 | for (size_t k = 1; k <= 80; k += 17) { |
| 14519 | GemmMicrokernelTester() |
| 14520 | .mr(2) |
| 14521 | .nr(8) |
| 14522 | .kr(8) |
| 14523 | .sr(1) |
| 14524 | .m(2) |
| 14525 | .n(n) |
| 14526 | .k(k) |
| 14527 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14528 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14529 | } |
| 14530 | } |
| 14531 | } |
| 14532 | |
| 14533 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, n_div_8_subtile) { |
| 14534 | TEST_REQUIRES_ARM_NEON; |
| 14535 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14536 | for (size_t k = 1; k <= 80; k += 17) { |
| 14537 | for (uint32_t m = 1; m <= 2; m++) { |
| 14538 | GemmMicrokernelTester() |
| 14539 | .mr(2) |
| 14540 | .nr(8) |
| 14541 | .kr(8) |
| 14542 | .sr(1) |
| 14543 | .m(m) |
| 14544 | .n(n) |
| 14545 | .k(k) |
| 14546 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14547 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14548 | } |
| 14549 | } |
| 14550 | } |
| 14551 | } |
| 14552 | |
| 14553 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, strided_cm_subtile) { |
| 14554 | TEST_REQUIRES_ARM_NEON; |
| 14555 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14556 | for (uint32_t n = 1; n <= 8; n++) { |
| 14557 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14558 | GemmMicrokernelTester() |
| 14559 | .mr(2) |
| 14560 | .nr(8) |
| 14561 | .kr(8) |
| 14562 | .sr(1) |
| 14563 | .m(m) |
| 14564 | .n(n) |
| 14565 | .k(k) |
| 14566 | .cm_stride(11) |
| 14567 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14568 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14569 | } |
| 14570 | } |
| 14571 | } |
| 14572 | } |
| 14573 | |
| 14574 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, qmin) { |
| 14575 | TEST_REQUIRES_ARM_NEON; |
| 14576 | GemmMicrokernelTester() |
| 14577 | .mr(2) |
| 14578 | .nr(8) |
| 14579 | .kr(8) |
| 14580 | .sr(1) |
| 14581 | .m(2) |
| 14582 | .n(8) |
| 14583 | .k(16) |
| 14584 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14585 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14586 | } |
| 14587 | |
| 14588 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, qmax) { |
| 14589 | TEST_REQUIRES_ARM_NEON; |
| 14590 | GemmMicrokernelTester() |
| 14591 | .mr(2) |
| 14592 | .nr(8) |
| 14593 | .kr(8) |
| 14594 | .sr(1) |
| 14595 | .m(2) |
| 14596 | .n(8) |
| 14597 | .k(16) |
| 14598 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14599 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14600 | } |
| 14601 | |
| 14602 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL_PRFM, strided_cm) { |
| 14603 | TEST_REQUIRES_ARM_NEON; |
| 14604 | GemmMicrokernelTester() |
| 14605 | .mr(2) |
| 14606 | .nr(8) |
| 14607 | .kr(8) |
| 14608 | .sr(1) |
| 14609 | .m(2) |
| 14610 | .n(8) |
| 14611 | .k(16) |
| 14612 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14613 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14614 | } |
| 14615 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 14616 | |
| 14617 | |
| 14618 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 14619 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_eq_8) { |
| 14620 | TEST_REQUIRES_ARM_NEON; |
| 14621 | GemmMicrokernelTester() |
| 14622 | .mr(2) |
| 14623 | .nr(8) |
| 14624 | .kr(8) |
| 14625 | .sr(1) |
| 14626 | .m(2) |
| 14627 | .n(8) |
| 14628 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14629 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14630 | } |
| 14631 | |
| 14632 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, strided_cn) { |
| 14633 | TEST_REQUIRES_ARM_NEON; |
| 14634 | GemmMicrokernelTester() |
| 14635 | .mr(2) |
| 14636 | .nr(8) |
| 14637 | .kr(8) |
| 14638 | .sr(1) |
| 14639 | .m(2) |
| 14640 | .n(8) |
| 14641 | .k(8) |
| 14642 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14643 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14644 | } |
| 14645 | |
| 14646 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_eq_8_strided_a) { |
| 14647 | TEST_REQUIRES_ARM_NEON; |
| 14648 | GemmMicrokernelTester() |
| 14649 | .mr(2) |
| 14650 | .nr(8) |
| 14651 | .kr(8) |
| 14652 | .sr(1) |
| 14653 | .m(2) |
| 14654 | .n(8) |
| 14655 | .k(8) |
| 14656 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14657 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14658 | } |
| 14659 | |
| 14660 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_eq_8_subtile) { |
| 14661 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14662 | for (uint32_t n = 1; n <= 8; n++) { |
| 14663 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14664 | GemmMicrokernelTester() |
| 14665 | .mr(2) |
| 14666 | .nr(8) |
| 14667 | .kr(8) |
| 14668 | .sr(1) |
| 14669 | .m(m) |
| 14670 | .n(n) |
| 14671 | .k(8) |
| 14672 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14673 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14674 | } |
| 14675 | } |
| 14676 | } |
| 14677 | |
| 14678 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_eq_8_subtile_m) { |
| 14679 | TEST_REQUIRES_ARM_NEON; |
| 14680 | for (uint32_t m = 1; m <= 2; m++) { |
| 14681 | GemmMicrokernelTester() |
| 14682 | .mr(2) |
| 14683 | .nr(8) |
| 14684 | .kr(8) |
| 14685 | .sr(1) |
| 14686 | .m(m) |
| 14687 | .n(8) |
| 14688 | .k(8) |
| 14689 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14690 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14691 | } |
| 14692 | } |
| 14693 | |
| 14694 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_eq_8_subtile_n) { |
| 14695 | TEST_REQUIRES_ARM_NEON; |
| 14696 | for (uint32_t n = 1; n <= 8; n++) { |
| 14697 | GemmMicrokernelTester() |
| 14698 | .mr(2) |
| 14699 | .nr(8) |
| 14700 | .kr(8) |
| 14701 | .sr(1) |
| 14702 | .m(2) |
| 14703 | .n(n) |
| 14704 | .k(8) |
| 14705 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14706 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14707 | } |
| 14708 | } |
| 14709 | |
| 14710 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_lt_8) { |
| 14711 | TEST_REQUIRES_ARM_NEON; |
| 14712 | for (size_t k = 1; k < 8; k++) { |
| 14713 | GemmMicrokernelTester() |
| 14714 | .mr(2) |
| 14715 | .nr(8) |
| 14716 | .kr(8) |
| 14717 | .sr(1) |
| 14718 | .m(2) |
| 14719 | .n(8) |
| 14720 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14721 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14722 | } |
| 14723 | } |
| 14724 | |
| 14725 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_lt_8_strided_a) { |
| 14726 | TEST_REQUIRES_ARM_NEON; |
| 14727 | for (size_t k = 1; k < 8; k++) { |
| 14728 | GemmMicrokernelTester() |
| 14729 | .mr(2) |
| 14730 | .nr(8) |
| 14731 | .kr(8) |
| 14732 | .sr(1) |
| 14733 | .m(2) |
| 14734 | .n(8) |
| 14735 | .k(k) |
| 14736 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14737 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14738 | } |
| 14739 | } |
| 14740 | |
| 14741 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_lt_8_subtile) { |
| 14742 | TEST_REQUIRES_ARM_NEON; |
| 14743 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14744 | for (uint32_t n = 1; n <= 8; n++) { |
| 14745 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14746 | GemmMicrokernelTester() |
| 14747 | .mr(2) |
| 14748 | .nr(8) |
| 14749 | .kr(8) |
| 14750 | .sr(1) |
| 14751 | .m(m) |
| 14752 | .n(n) |
| 14753 | .k(k) |
| 14754 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14755 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14756 | } |
| 14757 | } |
| 14758 | } |
| 14759 | } |
| 14760 | |
| 14761 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_gt_8) { |
| 14762 | TEST_REQUIRES_ARM_NEON; |
| 14763 | for (size_t k = 9; k < 16; k++) { |
| 14764 | GemmMicrokernelTester() |
| 14765 | .mr(2) |
| 14766 | .nr(8) |
| 14767 | .kr(8) |
| 14768 | .sr(1) |
| 14769 | .m(2) |
| 14770 | .n(8) |
| 14771 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14772 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14773 | } |
| 14774 | } |
| 14775 | |
| 14776 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_gt_8_strided_a) { |
| 14777 | TEST_REQUIRES_ARM_NEON; |
| 14778 | for (size_t k = 9; k < 16; k++) { |
| 14779 | GemmMicrokernelTester() |
| 14780 | .mr(2) |
| 14781 | .nr(8) |
| 14782 | .kr(8) |
| 14783 | .sr(1) |
| 14784 | .m(2) |
| 14785 | .n(8) |
| 14786 | .k(k) |
| 14787 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14788 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14789 | } |
| 14790 | } |
| 14791 | |
| 14792 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_gt_8_subtile) { |
| 14793 | TEST_REQUIRES_ARM_NEON; |
| 14794 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14795 | for (uint32_t n = 1; n <= 8; n++) { |
| 14796 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14797 | GemmMicrokernelTester() |
| 14798 | .mr(2) |
| 14799 | .nr(8) |
| 14800 | .kr(8) |
| 14801 | .sr(1) |
| 14802 | .m(m) |
| 14803 | .n(n) |
| 14804 | .k(k) |
| 14805 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14806 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14807 | } |
| 14808 | } |
| 14809 | } |
| 14810 | } |
| 14811 | |
| 14812 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_div_8) { |
| 14813 | TEST_REQUIRES_ARM_NEON; |
| 14814 | for (size_t k = 16; k <= 80; k += 8) { |
| 14815 | GemmMicrokernelTester() |
| 14816 | .mr(2) |
| 14817 | .nr(8) |
| 14818 | .kr(8) |
| 14819 | .sr(1) |
| 14820 | .m(2) |
| 14821 | .n(8) |
| 14822 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14823 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14824 | } |
| 14825 | } |
| 14826 | |
| 14827 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_div_8_strided_a) { |
| 14828 | TEST_REQUIRES_ARM_NEON; |
| 14829 | for (size_t k = 16; k <= 80; k += 8) { |
| 14830 | GemmMicrokernelTester() |
| 14831 | .mr(2) |
| 14832 | .nr(8) |
| 14833 | .kr(8) |
| 14834 | .sr(1) |
| 14835 | .m(2) |
| 14836 | .n(8) |
| 14837 | .k(k) |
| 14838 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14839 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14840 | } |
| 14841 | } |
| 14842 | |
| 14843 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, k_div_8_subtile) { |
| 14844 | TEST_REQUIRES_ARM_NEON; |
| 14845 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 14846 | for (uint32_t n = 1; n <= 8; n++) { |
| 14847 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14848 | GemmMicrokernelTester() |
| 14849 | .mr(2) |
| 14850 | .nr(8) |
| 14851 | .kr(8) |
| 14852 | .sr(1) |
| 14853 | .m(m) |
| 14854 | .n(n) |
| 14855 | .k(k) |
| 14856 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14857 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14858 | } |
| 14859 | } |
| 14860 | } |
| 14861 | } |
| 14862 | |
| 14863 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_gt_8) { |
| 14864 | TEST_REQUIRES_ARM_NEON; |
| 14865 | for (uint32_t n = 9; n < 16; n++) { |
| 14866 | for (size_t k = 1; k <= 40; k += 9) { |
| 14867 | GemmMicrokernelTester() |
| 14868 | .mr(2) |
| 14869 | .nr(8) |
| 14870 | .kr(8) |
| 14871 | .sr(1) |
| 14872 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 14873 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14874 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14875 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14876 | } |
| 14877 | } |
| 14878 | } |
| 14879 | |
| 14880 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_gt_8_strided_cn) { |
| 14881 | TEST_REQUIRES_ARM_NEON; |
| 14882 | for (uint32_t n = 9; n < 16; n++) { |
| 14883 | for (size_t k = 1; k <= 40; k += 9) { |
| 14884 | GemmMicrokernelTester() |
| 14885 | .mr(2) |
| 14886 | .nr(8) |
| 14887 | .kr(8) |
| 14888 | .sr(1) |
| 14889 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 14890 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14891 | .k(k) |
| 14892 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14893 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14894 | } |
| 14895 | } |
| 14896 | } |
| 14897 | |
| 14898 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_gt_8_strided_a) { |
| 14899 | TEST_REQUIRES_ARM_NEON; |
| 14900 | for (uint32_t n = 9; n < 16; n++) { |
| 14901 | for (size_t k = 1; k <= 40; k += 9) { |
| 14902 | GemmMicrokernelTester() |
| 14903 | .mr(2) |
| 14904 | .nr(8) |
| 14905 | .kr(8) |
| 14906 | .sr(1) |
| 14907 | .m(2) |
| 14908 | .n(n) |
| 14909 | .k(k) |
| 14910 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14911 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14912 | } |
| 14913 | } |
| 14914 | } |
| 14915 | |
| 14916 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_gt_8_subtile) { |
| 14917 | TEST_REQUIRES_ARM_NEON; |
| 14918 | for (uint32_t n = 9; n < 16; n++) { |
| 14919 | for (size_t k = 1; k <= 40; k += 9) { |
| 14920 | for (uint32_t m = 1; m <= 2; m++) { |
| 14921 | GemmMicrokernelTester() |
| 14922 | .mr(2) |
| 14923 | .nr(8) |
| 14924 | .kr(8) |
| 14925 | .sr(1) |
| 14926 | .m(m) |
| 14927 | .n(n) |
| 14928 | .k(k) |
| 14929 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14930 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14931 | } |
| 14932 | } |
| 14933 | } |
| 14934 | } |
| 14935 | |
| 14936 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_div_8) { |
| 14937 | TEST_REQUIRES_ARM_NEON; |
| 14938 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14939 | for (size_t k = 1; k <= 40; k += 9) { |
| 14940 | GemmMicrokernelTester() |
| 14941 | .mr(2) |
| 14942 | .nr(8) |
| 14943 | .kr(8) |
| 14944 | .sr(1) |
| 14945 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 14946 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14947 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14948 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14949 | } |
| 14950 | } |
| 14951 | } |
| 14952 | |
| 14953 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_div_8_strided_cn) { |
| 14954 | TEST_REQUIRES_ARM_NEON; |
| 14955 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14956 | for (size_t k = 1; k <= 40; k += 9) { |
| 14957 | GemmMicrokernelTester() |
| 14958 | .mr(2) |
| 14959 | .nr(8) |
| 14960 | .kr(8) |
| 14961 | .sr(1) |
| 14962 | .m(2) |
| 14963 | .n(n) |
| 14964 | .k(k) |
| 14965 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14966 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14967 | } |
| 14968 | } |
| 14969 | } |
| 14970 | |
| 14971 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_div_8_strided_a) { |
| 14972 | TEST_REQUIRES_ARM_NEON; |
| 14973 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14974 | for (size_t k = 1; k <= 40; k += 9) { |
| 14975 | GemmMicrokernelTester() |
| 14976 | .mr(2) |
| 14977 | .nr(8) |
| 14978 | .kr(8) |
| 14979 | .sr(1) |
| 14980 | .m(2) |
| 14981 | .n(n) |
| 14982 | .k(k) |
| 14983 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 14984 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 14985 | } |
| 14986 | } |
| 14987 | } |
| 14988 | |
| 14989 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, n_div_8_subtile) { |
| 14990 | TEST_REQUIRES_ARM_NEON; |
| 14991 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 14992 | for (size_t k = 1; k <= 40; k += 9) { |
| 14993 | for (uint32_t m = 1; m <= 2; m++) { |
| 14994 | GemmMicrokernelTester() |
| 14995 | .mr(2) |
| 14996 | .nr(8) |
| 14997 | .kr(8) |
| 14998 | .sr(1) |
| 14999 | .m(m) |
| 15000 | .n(n) |
| 15001 | .k(k) |
| 15002 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15003 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15004 | } |
| 15005 | } |
| 15006 | } |
| 15007 | } |
| 15008 | |
| 15009 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, strided_cm_subtile) { |
| 15010 | TEST_REQUIRES_ARM_NEON; |
| 15011 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15012 | for (uint32_t n = 1; n <= 8; n++) { |
| 15013 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15014 | GemmMicrokernelTester() |
| 15015 | .mr(2) |
| 15016 | .nr(8) |
| 15017 | .kr(8) |
| 15018 | .sr(1) |
| 15019 | .m(m) |
| 15020 | .n(n) |
| 15021 | .k(k) |
| 15022 | .cm_stride(11) |
| 15023 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15024 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15025 | } |
| 15026 | } |
| 15027 | } |
| 15028 | } |
| 15029 | |
| 15030 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, qmin) { |
| 15031 | TEST_REQUIRES_ARM_NEON; |
| 15032 | GemmMicrokernelTester() |
| 15033 | .mr(2) |
| 15034 | .nr(8) |
| 15035 | .kr(8) |
| 15036 | .sr(1) |
| 15037 | .m(2) |
| 15038 | .n(8) |
| 15039 | .k(8) |
| 15040 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15041 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15042 | } |
| 15043 | |
| 15044 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, qmax) { |
| 15045 | TEST_REQUIRES_ARM_NEON; |
| 15046 | GemmMicrokernelTester() |
| 15047 | .mr(2) |
| 15048 | .nr(8) |
| 15049 | .kr(8) |
| 15050 | .sr(1) |
| 15051 | .m(2) |
| 15052 | .n(8) |
| 15053 | .k(8) |
| 15054 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15055 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15056 | } |
| 15057 | |
| 15058 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MULL, strided_cm) { |
| 15059 | TEST_REQUIRES_ARM_NEON; |
| 15060 | GemmMicrokernelTester() |
| 15061 | .mr(2) |
| 15062 | .nr(8) |
| 15063 | .kr(8) |
| 15064 | .sr(1) |
| 15065 | .m(2) |
| 15066 | .n(8) |
| 15067 | .k(8) |
| 15068 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15069 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15070 | } |
| 15071 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 15072 | |
| 15073 | |
| 15074 | #if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 15075 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16) { |
| 15076 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15077 | GemmMicrokernelTester() |
| 15078 | .mr(4) |
| 15079 | .nr(16) |
| 15080 | .kr(4) |
| 15081 | .sr(1) |
| 15082 | .m(4) |
| 15083 | .n(16) |
| 15084 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15085 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15086 | } |
| 15087 | |
| 15088 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cn) { |
| 15089 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15090 | GemmMicrokernelTester() |
| 15091 | .mr(4) |
| 15092 | .nr(16) |
| 15093 | .kr(4) |
| 15094 | .sr(1) |
| 15095 | .m(4) |
| 15096 | .n(16) |
| 15097 | .k(16) |
| 15098 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15099 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15100 | } |
| 15101 | |
| 15102 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_strided_a) { |
| 15103 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15104 | GemmMicrokernelTester() |
| 15105 | .mr(4) |
| 15106 | .nr(16) |
| 15107 | .kr(4) |
| 15108 | .sr(1) |
| 15109 | .m(4) |
| 15110 | .n(16) |
| 15111 | .k(16) |
| 15112 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15113 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15114 | } |
| 15115 | |
| 15116 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile) { |
| 15117 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15118 | for (uint32_t n = 1; n <= 16; n++) { |
| 15119 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15120 | GemmMicrokernelTester() |
| 15121 | .mr(4) |
| 15122 | .nr(16) |
| 15123 | .kr(4) |
| 15124 | .sr(1) |
| 15125 | .m(m) |
| 15126 | .n(n) |
| 15127 | .k(16) |
| 15128 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15129 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15130 | } |
| 15131 | } |
| 15132 | } |
| 15133 | |
| 15134 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile_m) { |
| 15135 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15136 | for (uint32_t m = 1; m <= 4; m++) { |
| 15137 | GemmMicrokernelTester() |
| 15138 | .mr(4) |
| 15139 | .nr(16) |
| 15140 | .kr(4) |
| 15141 | .sr(1) |
| 15142 | .m(m) |
| 15143 | .n(16) |
| 15144 | .k(16) |
| 15145 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15146 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15147 | } |
| 15148 | } |
| 15149 | |
| 15150 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_eq_16_subtile_n) { |
| 15151 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15152 | for (uint32_t n = 1; n <= 16; n++) { |
| 15153 | GemmMicrokernelTester() |
| 15154 | .mr(4) |
| 15155 | .nr(16) |
| 15156 | .kr(4) |
| 15157 | .sr(1) |
| 15158 | .m(4) |
| 15159 | .n(n) |
| 15160 | .k(16) |
| 15161 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15162 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15163 | } |
| 15164 | } |
| 15165 | |
| 15166 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16) { |
| 15167 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15168 | for (size_t k = 1; k < 16; k++) { |
| 15169 | GemmMicrokernelTester() |
| 15170 | .mr(4) |
| 15171 | .nr(16) |
| 15172 | .kr(4) |
| 15173 | .sr(1) |
| 15174 | .m(4) |
| 15175 | .n(16) |
| 15176 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15177 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15178 | } |
| 15179 | } |
| 15180 | |
| 15181 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16_strided_a) { |
| 15182 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15183 | for (size_t k = 1; k < 16; k++) { |
| 15184 | GemmMicrokernelTester() |
| 15185 | .mr(4) |
| 15186 | .nr(16) |
| 15187 | .kr(4) |
| 15188 | .sr(1) |
| 15189 | .m(4) |
| 15190 | .n(16) |
| 15191 | .k(k) |
| 15192 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15193 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15194 | } |
| 15195 | } |
| 15196 | |
| 15197 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_lt_16_subtile) { |
| 15198 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15199 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15200 | for (uint32_t n = 1; n <= 16; n++) { |
| 15201 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15202 | GemmMicrokernelTester() |
| 15203 | .mr(4) |
| 15204 | .nr(16) |
| 15205 | .kr(4) |
| 15206 | .sr(1) |
| 15207 | .m(m) |
| 15208 | .n(n) |
| 15209 | .k(k) |
| 15210 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15211 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15212 | } |
| 15213 | } |
| 15214 | } |
| 15215 | } |
| 15216 | |
| 15217 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16) { |
| 15218 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15219 | for (size_t k = 17; k < 32; k++) { |
| 15220 | GemmMicrokernelTester() |
| 15221 | .mr(4) |
| 15222 | .nr(16) |
| 15223 | .kr(4) |
| 15224 | .sr(1) |
| 15225 | .m(4) |
| 15226 | .n(16) |
| 15227 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15228 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15229 | } |
| 15230 | } |
| 15231 | |
| 15232 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16_strided_a) { |
| 15233 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15234 | for (size_t k = 17; k < 32; k++) { |
| 15235 | GemmMicrokernelTester() |
| 15236 | .mr(4) |
| 15237 | .nr(16) |
| 15238 | .kr(4) |
| 15239 | .sr(1) |
| 15240 | .m(4) |
| 15241 | .n(16) |
| 15242 | .k(k) |
| 15243 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15244 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15245 | } |
| 15246 | } |
| 15247 | |
| 15248 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_gt_16_subtile) { |
| 15249 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15250 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15251 | for (uint32_t n = 1; n <= 16; n++) { |
| 15252 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15253 | GemmMicrokernelTester() |
| 15254 | .mr(4) |
| 15255 | .nr(16) |
| 15256 | .kr(4) |
| 15257 | .sr(1) |
| 15258 | .m(m) |
| 15259 | .n(n) |
| 15260 | .k(k) |
| 15261 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15262 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15263 | } |
| 15264 | } |
| 15265 | } |
| 15266 | } |
| 15267 | |
| 15268 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16) { |
| 15269 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15270 | for (size_t k = 32; k <= 160; k += 16) { |
| 15271 | GemmMicrokernelTester() |
| 15272 | .mr(4) |
| 15273 | .nr(16) |
| 15274 | .kr(4) |
| 15275 | .sr(1) |
| 15276 | .m(4) |
| 15277 | .n(16) |
| 15278 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15279 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15280 | } |
| 15281 | } |
| 15282 | |
| 15283 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16_strided_a) { |
| 15284 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15285 | for (size_t k = 32; k <= 160; k += 16) { |
| 15286 | GemmMicrokernelTester() |
| 15287 | .mr(4) |
| 15288 | .nr(16) |
| 15289 | .kr(4) |
| 15290 | .sr(1) |
| 15291 | .m(4) |
| 15292 | .n(16) |
| 15293 | .k(k) |
| 15294 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15295 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15296 | } |
| 15297 | } |
| 15298 | |
| 15299 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, k_div_16_subtile) { |
| 15300 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15301 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15302 | for (uint32_t n = 1; n <= 16; n++) { |
| 15303 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15304 | GemmMicrokernelTester() |
| 15305 | .mr(4) |
| 15306 | .nr(16) |
| 15307 | .kr(4) |
| 15308 | .sr(1) |
| 15309 | .m(m) |
| 15310 | .n(n) |
| 15311 | .k(k) |
| 15312 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15313 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15314 | } |
| 15315 | } |
| 15316 | } |
| 15317 | } |
| 15318 | |
| 15319 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16) { |
| 15320 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15321 | for (uint32_t n = 17; n < 32; n++) { |
| 15322 | for (size_t k = 1; k <= 80; k += 17) { |
| 15323 | GemmMicrokernelTester() |
| 15324 | .mr(4) |
| 15325 | .nr(16) |
| 15326 | .kr(4) |
| 15327 | .sr(1) |
| 15328 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 15329 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15330 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15331 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15332 | } |
| 15333 | } |
| 15334 | } |
| 15335 | |
| 15336 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_strided_cn) { |
| 15337 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15338 | for (uint32_t n = 17; n < 32; n++) { |
| 15339 | for (size_t k = 1; k <= 80; k += 17) { |
| 15340 | GemmMicrokernelTester() |
| 15341 | .mr(4) |
| 15342 | .nr(16) |
| 15343 | .kr(4) |
| 15344 | .sr(1) |
| 15345 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 15346 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15347 | .k(k) |
| 15348 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15349 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15350 | } |
| 15351 | } |
| 15352 | } |
| 15353 | |
| 15354 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_strided_a) { |
| 15355 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15356 | for (uint32_t n = 17; n < 32; n++) { |
| 15357 | for (size_t k = 1; k <= 80; k += 17) { |
| 15358 | GemmMicrokernelTester() |
| 15359 | .mr(4) |
| 15360 | .nr(16) |
| 15361 | .kr(4) |
| 15362 | .sr(1) |
| 15363 | .m(4) |
| 15364 | .n(n) |
| 15365 | .k(k) |
| 15366 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15367 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15368 | } |
| 15369 | } |
| 15370 | } |
| 15371 | |
| 15372 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_gt_16_subtile) { |
| 15373 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15374 | for (uint32_t n = 17; n < 32; n++) { |
| 15375 | for (size_t k = 1; k <= 80; k += 17) { |
| 15376 | for (uint32_t m = 1; m <= 4; m++) { |
| 15377 | GemmMicrokernelTester() |
| 15378 | .mr(4) |
| 15379 | .nr(16) |
| 15380 | .kr(4) |
| 15381 | .sr(1) |
| 15382 | .m(m) |
| 15383 | .n(n) |
| 15384 | .k(k) |
| 15385 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15386 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15387 | } |
| 15388 | } |
| 15389 | } |
| 15390 | } |
| 15391 | |
| 15392 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16) { |
| 15393 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15394 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 15395 | for (size_t k = 1; k <= 80; k += 17) { |
| 15396 | GemmMicrokernelTester() |
| 15397 | .mr(4) |
| 15398 | .nr(16) |
| 15399 | .kr(4) |
| 15400 | .sr(1) |
| 15401 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 15402 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15403 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15404 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15405 | } |
| 15406 | } |
| 15407 | } |
| 15408 | |
| 15409 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_strided_cn) { |
| 15410 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15411 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 15412 | for (size_t k = 1; k <= 80; k += 17) { |
| 15413 | GemmMicrokernelTester() |
| 15414 | .mr(4) |
| 15415 | .nr(16) |
| 15416 | .kr(4) |
| 15417 | .sr(1) |
| 15418 | .m(4) |
| 15419 | .n(n) |
| 15420 | .k(k) |
| 15421 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15422 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15423 | } |
| 15424 | } |
| 15425 | } |
| 15426 | |
| 15427 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_strided_a) { |
| 15428 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15429 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 15430 | for (size_t k = 1; k <= 80; k += 17) { |
| 15431 | GemmMicrokernelTester() |
| 15432 | .mr(4) |
| 15433 | .nr(16) |
| 15434 | .kr(4) |
| 15435 | .sr(1) |
| 15436 | .m(4) |
| 15437 | .n(n) |
| 15438 | .k(k) |
| 15439 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15440 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15441 | } |
| 15442 | } |
| 15443 | } |
| 15444 | |
| 15445 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, n_div_16_subtile) { |
| 15446 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15447 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 15448 | for (size_t k = 1; k <= 80; k += 17) { |
| 15449 | for (uint32_t m = 1; m <= 4; m++) { |
| 15450 | GemmMicrokernelTester() |
| 15451 | .mr(4) |
| 15452 | .nr(16) |
| 15453 | .kr(4) |
| 15454 | .sr(1) |
| 15455 | .m(m) |
| 15456 | .n(n) |
| 15457 | .k(k) |
| 15458 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15459 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15460 | } |
| 15461 | } |
| 15462 | } |
| 15463 | } |
| 15464 | |
| 15465 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cm_subtile) { |
| 15466 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15467 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15468 | for (uint32_t n = 1; n <= 16; n++) { |
| 15469 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15470 | GemmMicrokernelTester() |
| 15471 | .mr(4) |
| 15472 | .nr(16) |
| 15473 | .kr(4) |
| 15474 | .sr(1) |
| 15475 | .m(m) |
| 15476 | .n(n) |
| 15477 | .k(k) |
| 15478 | .cm_stride(19) |
| 15479 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15480 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15481 | } |
| 15482 | } |
| 15483 | } |
| 15484 | } |
| 15485 | |
| 15486 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, qmin) { |
| 15487 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15488 | GemmMicrokernelTester() |
| 15489 | .mr(4) |
| 15490 | .nr(16) |
| 15491 | .kr(4) |
| 15492 | .sr(1) |
| 15493 | .m(4) |
| 15494 | .n(16) |
| 15495 | .k(16) |
| 15496 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15497 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15498 | } |
| 15499 | |
| 15500 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, qmax) { |
| 15501 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15502 | GemmMicrokernelTester() |
| 15503 | .mr(4) |
| 15504 | .nr(16) |
| 15505 | .kr(4) |
| 15506 | .sr(1) |
| 15507 | .m(4) |
| 15508 | .n(16) |
| 15509 | .k(16) |
| 15510 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15511 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15512 | } |
| 15513 | |
| 15514 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_CORTEX_A55, strided_cm) { |
| 15515 | TEST_REQUIRES_ARM_NEON_DOT; |
| 15516 | GemmMicrokernelTester() |
| 15517 | .mr(4) |
| 15518 | .nr(16) |
| 15519 | .kr(4) |
| 15520 | .sr(1) |
| 15521 | .m(4) |
| 15522 | .n(16) |
| 15523 | .k(16) |
| 15524 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15525 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_cortex_a55, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15526 | } |
| 15527 | #endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY |
| 15528 | |
| 15529 | |
| 15530 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 15531 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8) { |
| 15532 | TEST_REQUIRES_ARM_NEON; |
| 15533 | GemmMicrokernelTester() |
| 15534 | .mr(2) |
| 15535 | .nr(8) |
| 15536 | .kr(8) |
| 15537 | .sr(1) |
| 15538 | .m(2) |
| 15539 | .n(8) |
| 15540 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15541 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15542 | } |
| 15543 | |
| 15544 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, strided_cn) { |
| 15545 | TEST_REQUIRES_ARM_NEON; |
| 15546 | GemmMicrokernelTester() |
| 15547 | .mr(2) |
| 15548 | .nr(8) |
| 15549 | .kr(8) |
| 15550 | .sr(1) |
| 15551 | .m(2) |
| 15552 | .n(8) |
| 15553 | .k(8) |
| 15554 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15555 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15556 | } |
| 15557 | |
| 15558 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8_strided_a) { |
| 15559 | TEST_REQUIRES_ARM_NEON; |
| 15560 | GemmMicrokernelTester() |
| 15561 | .mr(2) |
| 15562 | .nr(8) |
| 15563 | .kr(8) |
| 15564 | .sr(1) |
| 15565 | .m(2) |
| 15566 | .n(8) |
| 15567 | .k(8) |
| 15568 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15569 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15570 | } |
| 15571 | |
| 15572 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8_subtile) { |
| 15573 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15574 | for (uint32_t n = 1; n <= 8; n++) { |
| 15575 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15576 | GemmMicrokernelTester() |
| 15577 | .mr(2) |
| 15578 | .nr(8) |
| 15579 | .kr(8) |
| 15580 | .sr(1) |
| 15581 | .m(m) |
| 15582 | .n(n) |
| 15583 | .k(8) |
| 15584 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15585 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15586 | } |
| 15587 | } |
| 15588 | } |
| 15589 | |
| 15590 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8_subtile_m) { |
| 15591 | TEST_REQUIRES_ARM_NEON; |
| 15592 | for (uint32_t m = 1; m <= 2; m++) { |
| 15593 | GemmMicrokernelTester() |
| 15594 | .mr(2) |
| 15595 | .nr(8) |
| 15596 | .kr(8) |
| 15597 | .sr(1) |
| 15598 | .m(m) |
| 15599 | .n(8) |
| 15600 | .k(8) |
| 15601 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15602 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15603 | } |
| 15604 | } |
| 15605 | |
| 15606 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_eq_8_subtile_n) { |
| 15607 | TEST_REQUIRES_ARM_NEON; |
| 15608 | for (uint32_t n = 1; n <= 8; n++) { |
| 15609 | GemmMicrokernelTester() |
| 15610 | .mr(2) |
| 15611 | .nr(8) |
| 15612 | .kr(8) |
| 15613 | .sr(1) |
| 15614 | .m(2) |
| 15615 | .n(n) |
| 15616 | .k(8) |
| 15617 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15618 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15619 | } |
| 15620 | } |
| 15621 | |
| 15622 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_lt_8) { |
| 15623 | TEST_REQUIRES_ARM_NEON; |
| 15624 | for (size_t k = 1; k < 8; k++) { |
| 15625 | GemmMicrokernelTester() |
| 15626 | .mr(2) |
| 15627 | .nr(8) |
| 15628 | .kr(8) |
| 15629 | .sr(1) |
| 15630 | .m(2) |
| 15631 | .n(8) |
| 15632 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15633 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15634 | } |
| 15635 | } |
| 15636 | |
| 15637 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_lt_8_strided_a) { |
| 15638 | TEST_REQUIRES_ARM_NEON; |
| 15639 | for (size_t k = 1; k < 8; k++) { |
| 15640 | GemmMicrokernelTester() |
| 15641 | .mr(2) |
| 15642 | .nr(8) |
| 15643 | .kr(8) |
| 15644 | .sr(1) |
| 15645 | .m(2) |
| 15646 | .n(8) |
| 15647 | .k(k) |
| 15648 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15649 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15650 | } |
| 15651 | } |
| 15652 | |
| 15653 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_lt_8_subtile) { |
| 15654 | TEST_REQUIRES_ARM_NEON; |
| 15655 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15656 | for (uint32_t n = 1; n <= 8; n++) { |
| 15657 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15658 | GemmMicrokernelTester() |
| 15659 | .mr(2) |
| 15660 | .nr(8) |
| 15661 | .kr(8) |
| 15662 | .sr(1) |
| 15663 | .m(m) |
| 15664 | .n(n) |
| 15665 | .k(k) |
| 15666 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15667 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15668 | } |
| 15669 | } |
| 15670 | } |
| 15671 | } |
| 15672 | |
| 15673 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_gt_8) { |
| 15674 | TEST_REQUIRES_ARM_NEON; |
| 15675 | for (size_t k = 9; k < 16; k++) { |
| 15676 | GemmMicrokernelTester() |
| 15677 | .mr(2) |
| 15678 | .nr(8) |
| 15679 | .kr(8) |
| 15680 | .sr(1) |
| 15681 | .m(2) |
| 15682 | .n(8) |
| 15683 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15684 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15685 | } |
| 15686 | } |
| 15687 | |
| 15688 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_gt_8_strided_a) { |
| 15689 | TEST_REQUIRES_ARM_NEON; |
| 15690 | for (size_t k = 9; k < 16; k++) { |
| 15691 | GemmMicrokernelTester() |
| 15692 | .mr(2) |
| 15693 | .nr(8) |
| 15694 | .kr(8) |
| 15695 | .sr(1) |
| 15696 | .m(2) |
| 15697 | .n(8) |
| 15698 | .k(k) |
| 15699 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15700 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15701 | } |
| 15702 | } |
| 15703 | |
| 15704 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_gt_8_subtile) { |
| 15705 | TEST_REQUIRES_ARM_NEON; |
| 15706 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15707 | for (uint32_t n = 1; n <= 8; n++) { |
| 15708 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15709 | GemmMicrokernelTester() |
| 15710 | .mr(2) |
| 15711 | .nr(8) |
| 15712 | .kr(8) |
| 15713 | .sr(1) |
| 15714 | .m(m) |
| 15715 | .n(n) |
| 15716 | .k(k) |
| 15717 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15718 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15719 | } |
| 15720 | } |
| 15721 | } |
| 15722 | } |
| 15723 | |
| 15724 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_div_8) { |
| 15725 | TEST_REQUIRES_ARM_NEON; |
| 15726 | for (size_t k = 16; k <= 80; k += 8) { |
| 15727 | GemmMicrokernelTester() |
| 15728 | .mr(2) |
| 15729 | .nr(8) |
| 15730 | .kr(8) |
| 15731 | .sr(1) |
| 15732 | .m(2) |
| 15733 | .n(8) |
| 15734 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15735 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15736 | } |
| 15737 | } |
| 15738 | |
| 15739 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_div_8_strided_a) { |
| 15740 | TEST_REQUIRES_ARM_NEON; |
| 15741 | for (size_t k = 16; k <= 80; k += 8) { |
| 15742 | GemmMicrokernelTester() |
| 15743 | .mr(2) |
| 15744 | .nr(8) |
| 15745 | .kr(8) |
| 15746 | .sr(1) |
| 15747 | .m(2) |
| 15748 | .n(8) |
| 15749 | .k(k) |
| 15750 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15751 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15752 | } |
| 15753 | } |
| 15754 | |
| 15755 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, k_div_8_subtile) { |
| 15756 | TEST_REQUIRES_ARM_NEON; |
| 15757 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15758 | for (uint32_t n = 1; n <= 8; n++) { |
| 15759 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15760 | GemmMicrokernelTester() |
| 15761 | .mr(2) |
| 15762 | .nr(8) |
| 15763 | .kr(8) |
| 15764 | .sr(1) |
| 15765 | .m(m) |
| 15766 | .n(n) |
| 15767 | .k(k) |
| 15768 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15769 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15770 | } |
| 15771 | } |
| 15772 | } |
| 15773 | } |
| 15774 | |
| 15775 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8) { |
| 15776 | TEST_REQUIRES_ARM_NEON; |
| 15777 | for (uint32_t n = 9; n < 16; n++) { |
| 15778 | for (size_t k = 1; k <= 40; k += 9) { |
| 15779 | GemmMicrokernelTester() |
| 15780 | .mr(2) |
| 15781 | .nr(8) |
| 15782 | .kr(8) |
| 15783 | .sr(1) |
| 15784 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 15785 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15786 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15787 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15788 | } |
| 15789 | } |
| 15790 | } |
| 15791 | |
| 15792 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8_strided_cn) { |
| 15793 | TEST_REQUIRES_ARM_NEON; |
| 15794 | for (uint32_t n = 9; n < 16; n++) { |
| 15795 | for (size_t k = 1; k <= 40; k += 9) { |
| 15796 | GemmMicrokernelTester() |
| 15797 | .mr(2) |
| 15798 | .nr(8) |
| 15799 | .kr(8) |
| 15800 | .sr(1) |
| 15801 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 15802 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15803 | .k(k) |
| 15804 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15805 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15806 | } |
| 15807 | } |
| 15808 | } |
| 15809 | |
| 15810 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8_strided_a) { |
| 15811 | TEST_REQUIRES_ARM_NEON; |
| 15812 | for (uint32_t n = 9; n < 16; n++) { |
| 15813 | for (size_t k = 1; k <= 40; k += 9) { |
| 15814 | GemmMicrokernelTester() |
| 15815 | .mr(2) |
| 15816 | .nr(8) |
| 15817 | .kr(8) |
| 15818 | .sr(1) |
| 15819 | .m(2) |
| 15820 | .n(n) |
| 15821 | .k(k) |
| 15822 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15823 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15824 | } |
| 15825 | } |
| 15826 | } |
| 15827 | |
| 15828 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_gt_8_subtile) { |
| 15829 | TEST_REQUIRES_ARM_NEON; |
| 15830 | for (uint32_t n = 9; n < 16; n++) { |
| 15831 | for (size_t k = 1; k <= 40; k += 9) { |
| 15832 | for (uint32_t m = 1; m <= 2; m++) { |
| 15833 | GemmMicrokernelTester() |
| 15834 | .mr(2) |
| 15835 | .nr(8) |
| 15836 | .kr(8) |
| 15837 | .sr(1) |
| 15838 | .m(m) |
| 15839 | .n(n) |
| 15840 | .k(k) |
| 15841 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15842 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15843 | } |
| 15844 | } |
| 15845 | } |
| 15846 | } |
| 15847 | |
| 15848 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8) { |
| 15849 | TEST_REQUIRES_ARM_NEON; |
| 15850 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 15851 | for (size_t k = 1; k <= 40; k += 9) { |
| 15852 | GemmMicrokernelTester() |
| 15853 | .mr(2) |
| 15854 | .nr(8) |
| 15855 | .kr(8) |
| 15856 | .sr(1) |
| 15857 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 15858 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15859 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15860 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15861 | } |
| 15862 | } |
| 15863 | } |
| 15864 | |
| 15865 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8_strided_cn) { |
| 15866 | TEST_REQUIRES_ARM_NEON; |
| 15867 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 15868 | for (size_t k = 1; k <= 40; k += 9) { |
| 15869 | GemmMicrokernelTester() |
| 15870 | .mr(2) |
| 15871 | .nr(8) |
| 15872 | .kr(8) |
| 15873 | .sr(1) |
| 15874 | .m(2) |
| 15875 | .n(n) |
| 15876 | .k(k) |
| 15877 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15878 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15879 | } |
| 15880 | } |
| 15881 | } |
| 15882 | |
| 15883 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8_strided_a) { |
| 15884 | TEST_REQUIRES_ARM_NEON; |
| 15885 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 15886 | for (size_t k = 1; k <= 40; k += 9) { |
| 15887 | GemmMicrokernelTester() |
| 15888 | .mr(2) |
| 15889 | .nr(8) |
| 15890 | .kr(8) |
| 15891 | .sr(1) |
| 15892 | .m(2) |
| 15893 | .n(n) |
| 15894 | .k(k) |
| 15895 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15896 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15897 | } |
| 15898 | } |
| 15899 | } |
| 15900 | |
| 15901 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, n_div_8_subtile) { |
| 15902 | TEST_REQUIRES_ARM_NEON; |
| 15903 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 15904 | for (size_t k = 1; k <= 40; k += 9) { |
| 15905 | for (uint32_t m = 1; m <= 2; m++) { |
| 15906 | GemmMicrokernelTester() |
| 15907 | .mr(2) |
| 15908 | .nr(8) |
| 15909 | .kr(8) |
| 15910 | .sr(1) |
| 15911 | .m(m) |
| 15912 | .n(n) |
| 15913 | .k(k) |
| 15914 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15915 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15916 | } |
| 15917 | } |
| 15918 | } |
| 15919 | } |
| 15920 | |
| 15921 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, strided_cm_subtile) { |
| 15922 | TEST_REQUIRES_ARM_NEON; |
| 15923 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 15924 | for (uint32_t n = 1; n <= 8; n++) { |
| 15925 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15926 | GemmMicrokernelTester() |
| 15927 | .mr(2) |
| 15928 | .nr(8) |
| 15929 | .kr(8) |
| 15930 | .sr(1) |
| 15931 | .m(m) |
| 15932 | .n(n) |
| 15933 | .k(k) |
| 15934 | .cm_stride(11) |
| 15935 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15936 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15937 | } |
| 15938 | } |
| 15939 | } |
| 15940 | } |
| 15941 | |
| 15942 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, qmin) { |
| 15943 | TEST_REQUIRES_ARM_NEON; |
| 15944 | GemmMicrokernelTester() |
| 15945 | .mr(2) |
| 15946 | .nr(8) |
| 15947 | .kr(8) |
| 15948 | .sr(1) |
| 15949 | .m(2) |
| 15950 | .n(8) |
| 15951 | .k(8) |
| 15952 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15953 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15954 | } |
| 15955 | |
| 15956 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, qmax) { |
| 15957 | TEST_REQUIRES_ARM_NEON; |
| 15958 | GemmMicrokernelTester() |
| 15959 | .mr(2) |
| 15960 | .nr(8) |
| 15961 | .kr(8) |
| 15962 | .sr(1) |
| 15963 | .m(2) |
| 15964 | .n(8) |
| 15965 | .k(8) |
| 15966 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15967 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15968 | } |
| 15969 | |
| 15970 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8C8__NEON_MULL, strided_cm) { |
| 15971 | TEST_REQUIRES_ARM_NEON; |
| 15972 | GemmMicrokernelTester() |
| 15973 | .mr(2) |
| 15974 | .nr(8) |
| 15975 | .kr(8) |
| 15976 | .sr(1) |
| 15977 | .m(2) |
| 15978 | .n(8) |
| 15979 | .k(8) |
| 15980 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15981 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15982 | } |
| 15983 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 15984 | |
| 15985 | |
| 15986 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 15987 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8) { |
| 15988 | TEST_REQUIRES_ARM_NEON; |
| 15989 | GemmMicrokernelTester() |
| 15990 | .mr(3) |
| 15991 | .nr(8) |
| 15992 | .kr(8) |
| 15993 | .sr(1) |
| 15994 | .m(3) |
| 15995 | .n(8) |
| 15996 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 15997 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 15998 | } |
| 15999 | |
| 16000 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cn) { |
| 16001 | TEST_REQUIRES_ARM_NEON; |
| 16002 | GemmMicrokernelTester() |
| 16003 | .mr(3) |
| 16004 | .nr(8) |
| 16005 | .kr(8) |
| 16006 | .sr(1) |
| 16007 | .m(3) |
| 16008 | .n(8) |
| 16009 | .k(8) |
| 16010 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16011 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16012 | } |
| 16013 | |
| 16014 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_strided_a) { |
| 16015 | TEST_REQUIRES_ARM_NEON; |
| 16016 | GemmMicrokernelTester() |
| 16017 | .mr(3) |
| 16018 | .nr(8) |
| 16019 | .kr(8) |
| 16020 | .sr(1) |
| 16021 | .m(3) |
| 16022 | .n(8) |
| 16023 | .k(8) |
| 16024 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16025 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16026 | } |
| 16027 | |
| 16028 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile) { |
| 16029 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16030 | for (uint32_t n = 1; n <= 8; n++) { |
| 16031 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16032 | GemmMicrokernelTester() |
| 16033 | .mr(3) |
| 16034 | .nr(8) |
| 16035 | .kr(8) |
| 16036 | .sr(1) |
| 16037 | .m(m) |
| 16038 | .n(n) |
| 16039 | .k(8) |
| 16040 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16041 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16042 | } |
| 16043 | } |
| 16044 | } |
| 16045 | |
| 16046 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile_m) { |
| 16047 | TEST_REQUIRES_ARM_NEON; |
| 16048 | for (uint32_t m = 1; m <= 3; m++) { |
| 16049 | GemmMicrokernelTester() |
| 16050 | .mr(3) |
| 16051 | .nr(8) |
| 16052 | .kr(8) |
| 16053 | .sr(1) |
| 16054 | .m(m) |
| 16055 | .n(8) |
| 16056 | .k(8) |
| 16057 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16058 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16059 | } |
| 16060 | } |
| 16061 | |
| 16062 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile_n) { |
| 16063 | TEST_REQUIRES_ARM_NEON; |
| 16064 | for (uint32_t n = 1; n <= 8; n++) { |
| 16065 | GemmMicrokernelTester() |
| 16066 | .mr(3) |
| 16067 | .nr(8) |
| 16068 | .kr(8) |
| 16069 | .sr(1) |
| 16070 | .m(3) |
| 16071 | .n(n) |
| 16072 | .k(8) |
| 16073 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16074 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16075 | } |
| 16076 | } |
| 16077 | |
| 16078 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_lt_8) { |
| 16079 | TEST_REQUIRES_ARM_NEON; |
| 16080 | for (size_t k = 1; k < 8; k++) { |
| 16081 | GemmMicrokernelTester() |
| 16082 | .mr(3) |
| 16083 | .nr(8) |
| 16084 | .kr(8) |
| 16085 | .sr(1) |
| 16086 | .m(3) |
| 16087 | .n(8) |
| 16088 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16089 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16090 | } |
| 16091 | } |
| 16092 | |
| 16093 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_lt_8_strided_a) { |
| 16094 | TEST_REQUIRES_ARM_NEON; |
| 16095 | for (size_t k = 1; k < 8; k++) { |
| 16096 | GemmMicrokernelTester() |
| 16097 | .mr(3) |
| 16098 | .nr(8) |
| 16099 | .kr(8) |
| 16100 | .sr(1) |
| 16101 | .m(3) |
| 16102 | .n(8) |
| 16103 | .k(k) |
| 16104 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16105 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16106 | } |
| 16107 | } |
| 16108 | |
| 16109 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_lt_8_subtile) { |
| 16110 | TEST_REQUIRES_ARM_NEON; |
| 16111 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16112 | for (uint32_t n = 1; n <= 8; n++) { |
| 16113 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16114 | GemmMicrokernelTester() |
| 16115 | .mr(3) |
| 16116 | .nr(8) |
| 16117 | .kr(8) |
| 16118 | .sr(1) |
| 16119 | .m(m) |
| 16120 | .n(n) |
| 16121 | .k(k) |
| 16122 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16123 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16124 | } |
| 16125 | } |
| 16126 | } |
| 16127 | } |
| 16128 | |
| 16129 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_gt_8) { |
| 16130 | TEST_REQUIRES_ARM_NEON; |
| 16131 | for (size_t k = 9; k < 16; k++) { |
| 16132 | GemmMicrokernelTester() |
| 16133 | .mr(3) |
| 16134 | .nr(8) |
| 16135 | .kr(8) |
| 16136 | .sr(1) |
| 16137 | .m(3) |
| 16138 | .n(8) |
| 16139 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16140 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16141 | } |
| 16142 | } |
| 16143 | |
| 16144 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_gt_8_strided_a) { |
| 16145 | TEST_REQUIRES_ARM_NEON; |
| 16146 | for (size_t k = 9; k < 16; k++) { |
| 16147 | GemmMicrokernelTester() |
| 16148 | .mr(3) |
| 16149 | .nr(8) |
| 16150 | .kr(8) |
| 16151 | .sr(1) |
| 16152 | .m(3) |
| 16153 | .n(8) |
| 16154 | .k(k) |
| 16155 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16156 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16157 | } |
| 16158 | } |
| 16159 | |
| 16160 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_gt_8_subtile) { |
| 16161 | TEST_REQUIRES_ARM_NEON; |
| 16162 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16163 | for (uint32_t n = 1; n <= 8; n++) { |
| 16164 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16165 | GemmMicrokernelTester() |
| 16166 | .mr(3) |
| 16167 | .nr(8) |
| 16168 | .kr(8) |
| 16169 | .sr(1) |
| 16170 | .m(m) |
| 16171 | .n(n) |
| 16172 | .k(k) |
| 16173 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16174 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16175 | } |
| 16176 | } |
| 16177 | } |
| 16178 | } |
| 16179 | |
| 16180 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_div_8) { |
| 16181 | TEST_REQUIRES_ARM_NEON; |
| 16182 | for (size_t k = 16; k <= 80; k += 8) { |
| 16183 | GemmMicrokernelTester() |
| 16184 | .mr(3) |
| 16185 | .nr(8) |
| 16186 | .kr(8) |
| 16187 | .sr(1) |
| 16188 | .m(3) |
| 16189 | .n(8) |
| 16190 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16191 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16192 | } |
| 16193 | } |
| 16194 | |
| 16195 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_div_8_strided_a) { |
| 16196 | TEST_REQUIRES_ARM_NEON; |
| 16197 | for (size_t k = 16; k <= 80; k += 8) { |
| 16198 | GemmMicrokernelTester() |
| 16199 | .mr(3) |
| 16200 | .nr(8) |
| 16201 | .kr(8) |
| 16202 | .sr(1) |
| 16203 | .m(3) |
| 16204 | .n(8) |
| 16205 | .k(k) |
| 16206 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16207 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16208 | } |
| 16209 | } |
| 16210 | |
| 16211 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_div_8_subtile) { |
| 16212 | TEST_REQUIRES_ARM_NEON; |
| 16213 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16214 | for (uint32_t n = 1; n <= 8; n++) { |
| 16215 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16216 | GemmMicrokernelTester() |
| 16217 | .mr(3) |
| 16218 | .nr(8) |
| 16219 | .kr(8) |
| 16220 | .sr(1) |
| 16221 | .m(m) |
| 16222 | .n(n) |
| 16223 | .k(k) |
| 16224 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16225 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16226 | } |
| 16227 | } |
| 16228 | } |
| 16229 | } |
| 16230 | |
| 16231 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8) { |
| 16232 | TEST_REQUIRES_ARM_NEON; |
| 16233 | for (uint32_t n = 9; n < 16; n++) { |
| 16234 | for (size_t k = 1; k <= 40; k += 9) { |
| 16235 | GemmMicrokernelTester() |
| 16236 | .mr(3) |
| 16237 | .nr(8) |
| 16238 | .kr(8) |
| 16239 | .sr(1) |
| 16240 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 16241 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16242 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16243 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16244 | } |
| 16245 | } |
| 16246 | } |
| 16247 | |
| 16248 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_strided_cn) { |
| 16249 | TEST_REQUIRES_ARM_NEON; |
| 16250 | for (uint32_t n = 9; n < 16; n++) { |
| 16251 | for (size_t k = 1; k <= 40; k += 9) { |
| 16252 | GemmMicrokernelTester() |
| 16253 | .mr(3) |
| 16254 | .nr(8) |
| 16255 | .kr(8) |
| 16256 | .sr(1) |
| 16257 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 16258 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16259 | .k(k) |
| 16260 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16261 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16262 | } |
| 16263 | } |
| 16264 | } |
| 16265 | |
| 16266 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_strided_a) { |
| 16267 | TEST_REQUIRES_ARM_NEON; |
| 16268 | for (uint32_t n = 9; n < 16; n++) { |
| 16269 | for (size_t k = 1; k <= 40; k += 9) { |
| 16270 | GemmMicrokernelTester() |
| 16271 | .mr(3) |
| 16272 | .nr(8) |
| 16273 | .kr(8) |
| 16274 | .sr(1) |
| 16275 | .m(3) |
| 16276 | .n(n) |
| 16277 | .k(k) |
| 16278 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16279 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16280 | } |
| 16281 | } |
| 16282 | } |
| 16283 | |
| 16284 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_subtile) { |
| 16285 | TEST_REQUIRES_ARM_NEON; |
| 16286 | for (uint32_t n = 9; n < 16; n++) { |
| 16287 | for (size_t k = 1; k <= 40; k += 9) { |
| 16288 | for (uint32_t m = 1; m <= 3; m++) { |
| 16289 | GemmMicrokernelTester() |
| 16290 | .mr(3) |
| 16291 | .nr(8) |
| 16292 | .kr(8) |
| 16293 | .sr(1) |
| 16294 | .m(m) |
| 16295 | .n(n) |
| 16296 | .k(k) |
| 16297 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16298 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16299 | } |
| 16300 | } |
| 16301 | } |
| 16302 | } |
| 16303 | |
| 16304 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8) { |
| 16305 | TEST_REQUIRES_ARM_NEON; |
| 16306 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16307 | for (size_t k = 1; k <= 40; k += 9) { |
| 16308 | GemmMicrokernelTester() |
| 16309 | .mr(3) |
| 16310 | .nr(8) |
| 16311 | .kr(8) |
| 16312 | .sr(1) |
| 16313 | .m(3) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 16314 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16315 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16316 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16317 | } |
| 16318 | } |
| 16319 | } |
| 16320 | |
| 16321 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_strided_cn) { |
| 16322 | TEST_REQUIRES_ARM_NEON; |
| 16323 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16324 | for (size_t k = 1; k <= 40; k += 9) { |
| 16325 | GemmMicrokernelTester() |
| 16326 | .mr(3) |
| 16327 | .nr(8) |
| 16328 | .kr(8) |
| 16329 | .sr(1) |
| 16330 | .m(3) |
| 16331 | .n(n) |
| 16332 | .k(k) |
| 16333 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16334 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16335 | } |
| 16336 | } |
| 16337 | } |
| 16338 | |
| 16339 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_strided_a) { |
| 16340 | TEST_REQUIRES_ARM_NEON; |
| 16341 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16342 | for (size_t k = 1; k <= 40; k += 9) { |
| 16343 | GemmMicrokernelTester() |
| 16344 | .mr(3) |
| 16345 | .nr(8) |
| 16346 | .kr(8) |
| 16347 | .sr(1) |
| 16348 | .m(3) |
| 16349 | .n(n) |
| 16350 | .k(k) |
| 16351 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16352 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16353 | } |
| 16354 | } |
| 16355 | } |
| 16356 | |
| 16357 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_subtile) { |
| 16358 | TEST_REQUIRES_ARM_NEON; |
| 16359 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16360 | for (size_t k = 1; k <= 40; k += 9) { |
| 16361 | for (uint32_t m = 1; m <= 3; m++) { |
| 16362 | GemmMicrokernelTester() |
| 16363 | .mr(3) |
| 16364 | .nr(8) |
| 16365 | .kr(8) |
| 16366 | .sr(1) |
| 16367 | .m(m) |
| 16368 | .n(n) |
| 16369 | .k(k) |
| 16370 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16371 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16372 | } |
| 16373 | } |
| 16374 | } |
| 16375 | } |
| 16376 | |
| 16377 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cm_subtile) { |
| 16378 | TEST_REQUIRES_ARM_NEON; |
| 16379 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16380 | for (uint32_t n = 1; n <= 8; n++) { |
| 16381 | for (uint32_t m = 1; m <= 3; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16382 | GemmMicrokernelTester() |
| 16383 | .mr(3) |
| 16384 | .nr(8) |
| 16385 | .kr(8) |
| 16386 | .sr(1) |
| 16387 | .m(m) |
| 16388 | .n(n) |
| 16389 | .k(k) |
| 16390 | .cm_stride(11) |
| 16391 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16392 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16393 | } |
| 16394 | } |
| 16395 | } |
| 16396 | } |
| 16397 | |
| 16398 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, qmin) { |
| 16399 | TEST_REQUIRES_ARM_NEON; |
| 16400 | GemmMicrokernelTester() |
| 16401 | .mr(3) |
| 16402 | .nr(8) |
| 16403 | .kr(8) |
| 16404 | .sr(1) |
| 16405 | .m(3) |
| 16406 | .n(8) |
| 16407 | .k(8) |
| 16408 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16409 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16410 | } |
| 16411 | |
| 16412 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, qmax) { |
| 16413 | TEST_REQUIRES_ARM_NEON; |
| 16414 | GemmMicrokernelTester() |
| 16415 | .mr(3) |
| 16416 | .nr(8) |
| 16417 | .kr(8) |
| 16418 | .sr(1) |
| 16419 | .m(3) |
| 16420 | .n(8) |
| 16421 | .k(8) |
| 16422 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16423 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16424 | } |
| 16425 | |
| 16426 | TEST(QS8_GEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cm) { |
| 16427 | TEST_REQUIRES_ARM_NEON; |
| 16428 | GemmMicrokernelTester() |
| 16429 | .mr(3) |
| 16430 | .nr(8) |
| 16431 | .kr(8) |
| 16432 | .sr(1) |
| 16433 | .m(3) |
| 16434 | .n(8) |
| 16435 | .k(8) |
| 16436 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16437 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16438 | } |
| 16439 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 16440 | |
| 16441 | |
| 16442 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 16443 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_eq_8) { |
| 16444 | TEST_REQUIRES_ARM_NEON; |
| 16445 | GemmMicrokernelTester() |
| 16446 | .mr(4) |
| 16447 | .nr(8) |
| 16448 | .kr(8) |
| 16449 | .sr(1) |
| 16450 | .m(4) |
| 16451 | .n(8) |
| 16452 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16453 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16454 | } |
| 16455 | |
| 16456 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, strided_cn) { |
| 16457 | TEST_REQUIRES_ARM_NEON; |
| 16458 | GemmMicrokernelTester() |
| 16459 | .mr(4) |
| 16460 | .nr(8) |
| 16461 | .kr(8) |
| 16462 | .sr(1) |
| 16463 | .m(4) |
| 16464 | .n(8) |
| 16465 | .k(8) |
| 16466 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16467 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16468 | } |
| 16469 | |
| 16470 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_eq_8_strided_a) { |
| 16471 | TEST_REQUIRES_ARM_NEON; |
| 16472 | GemmMicrokernelTester() |
| 16473 | .mr(4) |
| 16474 | .nr(8) |
| 16475 | .kr(8) |
| 16476 | .sr(1) |
| 16477 | .m(4) |
| 16478 | .n(8) |
| 16479 | .k(8) |
| 16480 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16481 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16482 | } |
| 16483 | |
| 16484 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_eq_8_subtile) { |
| 16485 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16486 | for (uint32_t n = 1; n <= 8; n++) { |
| 16487 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16488 | GemmMicrokernelTester() |
| 16489 | .mr(4) |
| 16490 | .nr(8) |
| 16491 | .kr(8) |
| 16492 | .sr(1) |
| 16493 | .m(m) |
| 16494 | .n(n) |
| 16495 | .k(8) |
| 16496 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16497 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16498 | } |
| 16499 | } |
| 16500 | } |
| 16501 | |
| 16502 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_eq_8_subtile_m) { |
| 16503 | TEST_REQUIRES_ARM_NEON; |
| 16504 | for (uint32_t m = 1; m <= 4; m++) { |
| 16505 | GemmMicrokernelTester() |
| 16506 | .mr(4) |
| 16507 | .nr(8) |
| 16508 | .kr(8) |
| 16509 | .sr(1) |
| 16510 | .m(m) |
| 16511 | .n(8) |
| 16512 | .k(8) |
| 16513 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16514 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16515 | } |
| 16516 | } |
| 16517 | |
| 16518 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_eq_8_subtile_n) { |
| 16519 | TEST_REQUIRES_ARM_NEON; |
| 16520 | for (uint32_t n = 1; n <= 8; n++) { |
| 16521 | GemmMicrokernelTester() |
| 16522 | .mr(4) |
| 16523 | .nr(8) |
| 16524 | .kr(8) |
| 16525 | .sr(1) |
| 16526 | .m(4) |
| 16527 | .n(n) |
| 16528 | .k(8) |
| 16529 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16530 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16531 | } |
| 16532 | } |
| 16533 | |
| 16534 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_lt_8) { |
| 16535 | TEST_REQUIRES_ARM_NEON; |
| 16536 | for (size_t k = 1; k < 8; k++) { |
| 16537 | GemmMicrokernelTester() |
| 16538 | .mr(4) |
| 16539 | .nr(8) |
| 16540 | .kr(8) |
| 16541 | .sr(1) |
| 16542 | .m(4) |
| 16543 | .n(8) |
| 16544 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16545 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16546 | } |
| 16547 | } |
| 16548 | |
| 16549 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_lt_8_strided_a) { |
| 16550 | TEST_REQUIRES_ARM_NEON; |
| 16551 | for (size_t k = 1; k < 8; k++) { |
| 16552 | GemmMicrokernelTester() |
| 16553 | .mr(4) |
| 16554 | .nr(8) |
| 16555 | .kr(8) |
| 16556 | .sr(1) |
| 16557 | .m(4) |
| 16558 | .n(8) |
| 16559 | .k(k) |
| 16560 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16561 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16562 | } |
| 16563 | } |
| 16564 | |
| 16565 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_lt_8_subtile) { |
| 16566 | TEST_REQUIRES_ARM_NEON; |
| 16567 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16568 | for (uint32_t n = 1; n <= 8; n++) { |
| 16569 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16570 | GemmMicrokernelTester() |
| 16571 | .mr(4) |
| 16572 | .nr(8) |
| 16573 | .kr(8) |
| 16574 | .sr(1) |
| 16575 | .m(m) |
| 16576 | .n(n) |
| 16577 | .k(k) |
| 16578 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16579 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16580 | } |
| 16581 | } |
| 16582 | } |
| 16583 | } |
| 16584 | |
| 16585 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_gt_8) { |
| 16586 | TEST_REQUIRES_ARM_NEON; |
| 16587 | for (size_t k = 9; k < 16; k++) { |
| 16588 | GemmMicrokernelTester() |
| 16589 | .mr(4) |
| 16590 | .nr(8) |
| 16591 | .kr(8) |
| 16592 | .sr(1) |
| 16593 | .m(4) |
| 16594 | .n(8) |
| 16595 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16596 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16597 | } |
| 16598 | } |
| 16599 | |
| 16600 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_gt_8_strided_a) { |
| 16601 | TEST_REQUIRES_ARM_NEON; |
| 16602 | for (size_t k = 9; k < 16; k++) { |
| 16603 | GemmMicrokernelTester() |
| 16604 | .mr(4) |
| 16605 | .nr(8) |
| 16606 | .kr(8) |
| 16607 | .sr(1) |
| 16608 | .m(4) |
| 16609 | .n(8) |
| 16610 | .k(k) |
| 16611 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16612 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16613 | } |
| 16614 | } |
| 16615 | |
| 16616 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_gt_8_subtile) { |
| 16617 | TEST_REQUIRES_ARM_NEON; |
| 16618 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16619 | for (uint32_t n = 1; n <= 8; n++) { |
| 16620 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16621 | GemmMicrokernelTester() |
| 16622 | .mr(4) |
| 16623 | .nr(8) |
| 16624 | .kr(8) |
| 16625 | .sr(1) |
| 16626 | .m(m) |
| 16627 | .n(n) |
| 16628 | .k(k) |
| 16629 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16630 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16631 | } |
| 16632 | } |
| 16633 | } |
| 16634 | } |
| 16635 | |
| 16636 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_div_8) { |
| 16637 | TEST_REQUIRES_ARM_NEON; |
| 16638 | for (size_t k = 16; k <= 80; k += 8) { |
| 16639 | GemmMicrokernelTester() |
| 16640 | .mr(4) |
| 16641 | .nr(8) |
| 16642 | .kr(8) |
| 16643 | .sr(1) |
| 16644 | .m(4) |
| 16645 | .n(8) |
| 16646 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16647 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16648 | } |
| 16649 | } |
| 16650 | |
| 16651 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_div_8_strided_a) { |
| 16652 | TEST_REQUIRES_ARM_NEON; |
| 16653 | for (size_t k = 16; k <= 80; k += 8) { |
| 16654 | GemmMicrokernelTester() |
| 16655 | .mr(4) |
| 16656 | .nr(8) |
| 16657 | .kr(8) |
| 16658 | .sr(1) |
| 16659 | .m(4) |
| 16660 | .n(8) |
| 16661 | .k(k) |
| 16662 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16663 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16664 | } |
| 16665 | } |
| 16666 | |
| 16667 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, k_div_8_subtile) { |
| 16668 | TEST_REQUIRES_ARM_NEON; |
| 16669 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16670 | for (uint32_t n = 1; n <= 8; n++) { |
| 16671 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16672 | GemmMicrokernelTester() |
| 16673 | .mr(4) |
| 16674 | .nr(8) |
| 16675 | .kr(8) |
| 16676 | .sr(1) |
| 16677 | .m(m) |
| 16678 | .n(n) |
| 16679 | .k(k) |
| 16680 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16681 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16682 | } |
| 16683 | } |
| 16684 | } |
| 16685 | } |
| 16686 | |
| 16687 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_gt_8) { |
| 16688 | TEST_REQUIRES_ARM_NEON; |
| 16689 | for (uint32_t n = 9; n < 16; n++) { |
| 16690 | for (size_t k = 1; k <= 40; k += 9) { |
| 16691 | GemmMicrokernelTester() |
| 16692 | .mr(4) |
| 16693 | .nr(8) |
| 16694 | .kr(8) |
| 16695 | .sr(1) |
| 16696 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 16697 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16698 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16699 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16700 | } |
| 16701 | } |
| 16702 | } |
| 16703 | |
| 16704 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_gt_8_strided_cn) { |
| 16705 | TEST_REQUIRES_ARM_NEON; |
| 16706 | for (uint32_t n = 9; n < 16; n++) { |
| 16707 | for (size_t k = 1; k <= 40; k += 9) { |
| 16708 | GemmMicrokernelTester() |
| 16709 | .mr(4) |
| 16710 | .nr(8) |
| 16711 | .kr(8) |
| 16712 | .sr(1) |
| 16713 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 16714 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16715 | .k(k) |
| 16716 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16717 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16718 | } |
| 16719 | } |
| 16720 | } |
| 16721 | |
| 16722 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_gt_8_strided_a) { |
| 16723 | TEST_REQUIRES_ARM_NEON; |
| 16724 | for (uint32_t n = 9; n < 16; n++) { |
| 16725 | for (size_t k = 1; k <= 40; k += 9) { |
| 16726 | GemmMicrokernelTester() |
| 16727 | .mr(4) |
| 16728 | .nr(8) |
| 16729 | .kr(8) |
| 16730 | .sr(1) |
| 16731 | .m(4) |
| 16732 | .n(n) |
| 16733 | .k(k) |
| 16734 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16735 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16736 | } |
| 16737 | } |
| 16738 | } |
| 16739 | |
| 16740 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_gt_8_subtile) { |
| 16741 | TEST_REQUIRES_ARM_NEON; |
| 16742 | for (uint32_t n = 9; n < 16; n++) { |
| 16743 | for (size_t k = 1; k <= 40; k += 9) { |
| 16744 | for (uint32_t m = 1; m <= 4; m++) { |
| 16745 | GemmMicrokernelTester() |
| 16746 | .mr(4) |
| 16747 | .nr(8) |
| 16748 | .kr(8) |
| 16749 | .sr(1) |
| 16750 | .m(m) |
| 16751 | .n(n) |
| 16752 | .k(k) |
| 16753 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16754 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16755 | } |
| 16756 | } |
| 16757 | } |
| 16758 | } |
| 16759 | |
| 16760 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_div_8) { |
| 16761 | TEST_REQUIRES_ARM_NEON; |
| 16762 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16763 | for (size_t k = 1; k <= 40; k += 9) { |
| 16764 | GemmMicrokernelTester() |
| 16765 | .mr(4) |
| 16766 | .nr(8) |
| 16767 | .kr(8) |
| 16768 | .sr(1) |
| 16769 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 16770 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16771 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16772 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16773 | } |
| 16774 | } |
| 16775 | } |
| 16776 | |
| 16777 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_div_8_strided_cn) { |
| 16778 | TEST_REQUIRES_ARM_NEON; |
| 16779 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16780 | for (size_t k = 1; k <= 40; k += 9) { |
| 16781 | GemmMicrokernelTester() |
| 16782 | .mr(4) |
| 16783 | .nr(8) |
| 16784 | .kr(8) |
| 16785 | .sr(1) |
| 16786 | .m(4) |
| 16787 | .n(n) |
| 16788 | .k(k) |
| 16789 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16790 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16791 | } |
| 16792 | } |
| 16793 | } |
| 16794 | |
| 16795 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_div_8_strided_a) { |
| 16796 | TEST_REQUIRES_ARM_NEON; |
| 16797 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16798 | for (size_t k = 1; k <= 40; k += 9) { |
| 16799 | GemmMicrokernelTester() |
| 16800 | .mr(4) |
| 16801 | .nr(8) |
| 16802 | .kr(8) |
| 16803 | .sr(1) |
| 16804 | .m(4) |
| 16805 | .n(n) |
| 16806 | .k(k) |
| 16807 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16808 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16809 | } |
| 16810 | } |
| 16811 | } |
| 16812 | |
| 16813 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, n_div_8_subtile) { |
| 16814 | TEST_REQUIRES_ARM_NEON; |
| 16815 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 16816 | for (size_t k = 1; k <= 40; k += 9) { |
| 16817 | for (uint32_t m = 1; m <= 4; m++) { |
| 16818 | GemmMicrokernelTester() |
| 16819 | .mr(4) |
| 16820 | .nr(8) |
| 16821 | .kr(8) |
| 16822 | .sr(1) |
| 16823 | .m(m) |
| 16824 | .n(n) |
| 16825 | .k(k) |
| 16826 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16827 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16828 | } |
| 16829 | } |
| 16830 | } |
| 16831 | } |
| 16832 | |
| 16833 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, strided_cm_subtile) { |
| 16834 | TEST_REQUIRES_ARM_NEON; |
| 16835 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16836 | for (uint32_t n = 1; n <= 8; n++) { |
| 16837 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16838 | GemmMicrokernelTester() |
| 16839 | .mr(4) |
| 16840 | .nr(8) |
| 16841 | .kr(8) |
| 16842 | .sr(1) |
| 16843 | .m(m) |
| 16844 | .n(n) |
| 16845 | .k(k) |
| 16846 | .cm_stride(11) |
| 16847 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16848 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16849 | } |
| 16850 | } |
| 16851 | } |
| 16852 | } |
| 16853 | |
| 16854 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, qmin) { |
| 16855 | TEST_REQUIRES_ARM_NEON; |
| 16856 | GemmMicrokernelTester() |
| 16857 | .mr(4) |
| 16858 | .nr(8) |
| 16859 | .kr(8) |
| 16860 | .sr(1) |
| 16861 | .m(4) |
| 16862 | .n(8) |
| 16863 | .k(8) |
| 16864 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16865 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16866 | } |
| 16867 | |
| 16868 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, qmax) { |
| 16869 | TEST_REQUIRES_ARM_NEON; |
| 16870 | GemmMicrokernelTester() |
| 16871 | .mr(4) |
| 16872 | .nr(8) |
| 16873 | .kr(8) |
| 16874 | .sr(1) |
| 16875 | .m(4) |
| 16876 | .n(8) |
| 16877 | .k(8) |
| 16878 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16879 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16880 | } |
| 16881 | |
| 16882 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8C8__NEON_MULL, strided_cm) { |
| 16883 | TEST_REQUIRES_ARM_NEON; |
| 16884 | GemmMicrokernelTester() |
| 16885 | .mr(4) |
| 16886 | .nr(8) |
| 16887 | .kr(8) |
| 16888 | .sr(1) |
| 16889 | .m(4) |
| 16890 | .n(8) |
| 16891 | .k(8) |
| 16892 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16893 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16894 | } |
| 16895 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 16896 | |
| 16897 | |
| 16898 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 16899 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_eq_16) { |
| 16900 | TEST_REQUIRES_ARM_NEON; |
| 16901 | GemmMicrokernelTester() |
| 16902 | .mr(1) |
| 16903 | .nr(8) |
| 16904 | .kr(8) |
| 16905 | .sr(1) |
| 16906 | .m(1) |
| 16907 | .n(8) |
| 16908 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16909 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16910 | } |
| 16911 | |
| 16912 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, strided_cn) { |
| 16913 | TEST_REQUIRES_ARM_NEON; |
| 16914 | GemmMicrokernelTester() |
| 16915 | .mr(1) |
| 16916 | .nr(8) |
| 16917 | .kr(8) |
| 16918 | .sr(1) |
| 16919 | .m(1) |
| 16920 | .n(8) |
| 16921 | .k(16) |
| 16922 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16923 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16924 | } |
| 16925 | |
| 16926 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_eq_16_strided_a) { |
| 16927 | TEST_REQUIRES_ARM_NEON; |
| 16928 | GemmMicrokernelTester() |
| 16929 | .mr(1) |
| 16930 | .nr(8) |
| 16931 | .kr(8) |
| 16932 | .sr(1) |
| 16933 | .m(1) |
| 16934 | .n(8) |
| 16935 | .k(16) |
| 16936 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16937 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16938 | } |
| 16939 | |
| 16940 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_eq_16_subtile) { |
| 16941 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 16942 | for (uint32_t n = 1; n <= 8; n++) { |
| 16943 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16944 | GemmMicrokernelTester() |
| 16945 | .mr(1) |
| 16946 | .nr(8) |
| 16947 | .kr(8) |
| 16948 | .sr(1) |
| 16949 | .m(m) |
| 16950 | .n(n) |
| 16951 | .k(16) |
| 16952 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16953 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16954 | } |
| 16955 | } |
| 16956 | } |
| 16957 | |
| 16958 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_eq_16_subtile_m) { |
| 16959 | TEST_REQUIRES_ARM_NEON; |
| 16960 | for (uint32_t m = 1; m <= 1; m++) { |
| 16961 | GemmMicrokernelTester() |
| 16962 | .mr(1) |
| 16963 | .nr(8) |
| 16964 | .kr(8) |
| 16965 | .sr(1) |
| 16966 | .m(m) |
| 16967 | .n(8) |
| 16968 | .k(16) |
| 16969 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16970 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16971 | } |
| 16972 | } |
| 16973 | |
| 16974 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_eq_16_subtile_n) { |
| 16975 | TEST_REQUIRES_ARM_NEON; |
| 16976 | for (uint32_t n = 1; n <= 8; n++) { |
| 16977 | GemmMicrokernelTester() |
| 16978 | .mr(1) |
| 16979 | .nr(8) |
| 16980 | .kr(8) |
| 16981 | .sr(1) |
| 16982 | .m(1) |
| 16983 | .n(n) |
| 16984 | .k(16) |
| 16985 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 16986 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 16987 | } |
| 16988 | } |
| 16989 | |
| 16990 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_lt_16) { |
| 16991 | TEST_REQUIRES_ARM_NEON; |
| 16992 | for (size_t k = 1; k < 16; k++) { |
| 16993 | GemmMicrokernelTester() |
| 16994 | .mr(1) |
| 16995 | .nr(8) |
| 16996 | .kr(8) |
| 16997 | .sr(1) |
| 16998 | .m(1) |
| 16999 | .n(8) |
| 17000 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17001 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17002 | } |
| 17003 | } |
| 17004 | |
| 17005 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_lt_16_strided_a) { |
| 17006 | TEST_REQUIRES_ARM_NEON; |
| 17007 | for (size_t k = 1; k < 16; k++) { |
| 17008 | GemmMicrokernelTester() |
| 17009 | .mr(1) |
| 17010 | .nr(8) |
| 17011 | .kr(8) |
| 17012 | .sr(1) |
| 17013 | .m(1) |
| 17014 | .n(8) |
| 17015 | .k(k) |
| 17016 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17017 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17018 | } |
| 17019 | } |
| 17020 | |
| 17021 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_lt_16_subtile) { |
| 17022 | TEST_REQUIRES_ARM_NEON; |
| 17023 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17024 | for (uint32_t n = 1; n <= 8; n++) { |
| 17025 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17026 | GemmMicrokernelTester() |
| 17027 | .mr(1) |
| 17028 | .nr(8) |
| 17029 | .kr(8) |
| 17030 | .sr(1) |
| 17031 | .m(m) |
| 17032 | .n(n) |
| 17033 | .k(k) |
| 17034 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17035 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17036 | } |
| 17037 | } |
| 17038 | } |
| 17039 | } |
| 17040 | |
| 17041 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_gt_16) { |
| 17042 | TEST_REQUIRES_ARM_NEON; |
| 17043 | for (size_t k = 17; k < 32; k++) { |
| 17044 | GemmMicrokernelTester() |
| 17045 | .mr(1) |
| 17046 | .nr(8) |
| 17047 | .kr(8) |
| 17048 | .sr(1) |
| 17049 | .m(1) |
| 17050 | .n(8) |
| 17051 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17052 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17053 | } |
| 17054 | } |
| 17055 | |
| 17056 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_gt_16_strided_a) { |
| 17057 | TEST_REQUIRES_ARM_NEON; |
| 17058 | for (size_t k = 17; k < 32; k++) { |
| 17059 | GemmMicrokernelTester() |
| 17060 | .mr(1) |
| 17061 | .nr(8) |
| 17062 | .kr(8) |
| 17063 | .sr(1) |
| 17064 | .m(1) |
| 17065 | .n(8) |
| 17066 | .k(k) |
| 17067 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17068 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17069 | } |
| 17070 | } |
| 17071 | |
| 17072 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_gt_16_subtile) { |
| 17073 | TEST_REQUIRES_ARM_NEON; |
| 17074 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17075 | for (uint32_t n = 1; n <= 8; n++) { |
| 17076 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17077 | GemmMicrokernelTester() |
| 17078 | .mr(1) |
| 17079 | .nr(8) |
| 17080 | .kr(8) |
| 17081 | .sr(1) |
| 17082 | .m(m) |
| 17083 | .n(n) |
| 17084 | .k(k) |
| 17085 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17086 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17087 | } |
| 17088 | } |
| 17089 | } |
| 17090 | } |
| 17091 | |
| 17092 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_div_16) { |
| 17093 | TEST_REQUIRES_ARM_NEON; |
| 17094 | for (size_t k = 32; k <= 160; k += 16) { |
| 17095 | GemmMicrokernelTester() |
| 17096 | .mr(1) |
| 17097 | .nr(8) |
| 17098 | .kr(8) |
| 17099 | .sr(1) |
| 17100 | .m(1) |
| 17101 | .n(8) |
| 17102 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17103 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17104 | } |
| 17105 | } |
| 17106 | |
| 17107 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_div_16_strided_a) { |
| 17108 | TEST_REQUIRES_ARM_NEON; |
| 17109 | for (size_t k = 32; k <= 160; k += 16) { |
| 17110 | GemmMicrokernelTester() |
| 17111 | .mr(1) |
| 17112 | .nr(8) |
| 17113 | .kr(8) |
| 17114 | .sr(1) |
| 17115 | .m(1) |
| 17116 | .n(8) |
| 17117 | .k(k) |
| 17118 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17119 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17120 | } |
| 17121 | } |
| 17122 | |
| 17123 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, k_div_16_subtile) { |
| 17124 | TEST_REQUIRES_ARM_NEON; |
| 17125 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17126 | for (uint32_t n = 1; n <= 8; n++) { |
| 17127 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17128 | GemmMicrokernelTester() |
| 17129 | .mr(1) |
| 17130 | .nr(8) |
| 17131 | .kr(8) |
| 17132 | .sr(1) |
| 17133 | .m(m) |
| 17134 | .n(n) |
| 17135 | .k(k) |
| 17136 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17137 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17138 | } |
| 17139 | } |
| 17140 | } |
| 17141 | } |
| 17142 | |
| 17143 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_gt_8) { |
| 17144 | TEST_REQUIRES_ARM_NEON; |
| 17145 | for (uint32_t n = 9; n < 16; n++) { |
| 17146 | for (size_t k = 1; k <= 80; k += 17) { |
| 17147 | GemmMicrokernelTester() |
| 17148 | .mr(1) |
| 17149 | .nr(8) |
| 17150 | .kr(8) |
| 17151 | .sr(1) |
| 17152 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 17153 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17154 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17155 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17156 | } |
| 17157 | } |
| 17158 | } |
| 17159 | |
| 17160 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_gt_8_strided_cn) { |
| 17161 | TEST_REQUIRES_ARM_NEON; |
| 17162 | for (uint32_t n = 9; n < 16; n++) { |
| 17163 | for (size_t k = 1; k <= 80; k += 17) { |
| 17164 | GemmMicrokernelTester() |
| 17165 | .mr(1) |
| 17166 | .nr(8) |
| 17167 | .kr(8) |
| 17168 | .sr(1) |
| 17169 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 17170 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17171 | .k(k) |
| 17172 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17173 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17174 | } |
| 17175 | } |
| 17176 | } |
| 17177 | |
| 17178 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_gt_8_strided_a) { |
| 17179 | TEST_REQUIRES_ARM_NEON; |
| 17180 | for (uint32_t n = 9; n < 16; n++) { |
| 17181 | for (size_t k = 1; k <= 80; k += 17) { |
| 17182 | GemmMicrokernelTester() |
| 17183 | .mr(1) |
| 17184 | .nr(8) |
| 17185 | .kr(8) |
| 17186 | .sr(1) |
| 17187 | .m(1) |
| 17188 | .n(n) |
| 17189 | .k(k) |
| 17190 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17191 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17192 | } |
| 17193 | } |
| 17194 | } |
| 17195 | |
| 17196 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_gt_8_subtile) { |
| 17197 | TEST_REQUIRES_ARM_NEON; |
| 17198 | for (uint32_t n = 9; n < 16; n++) { |
| 17199 | for (size_t k = 1; k <= 80; k += 17) { |
| 17200 | for (uint32_t m = 1; m <= 1; m++) { |
| 17201 | GemmMicrokernelTester() |
| 17202 | .mr(1) |
| 17203 | .nr(8) |
| 17204 | .kr(8) |
| 17205 | .sr(1) |
| 17206 | .m(m) |
| 17207 | .n(n) |
| 17208 | .k(k) |
| 17209 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17210 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17211 | } |
| 17212 | } |
| 17213 | } |
| 17214 | } |
| 17215 | |
| 17216 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_div_8) { |
| 17217 | TEST_REQUIRES_ARM_NEON; |
| 17218 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 17219 | for (size_t k = 1; k <= 80; k += 17) { |
| 17220 | GemmMicrokernelTester() |
| 17221 | .mr(1) |
| 17222 | .nr(8) |
| 17223 | .kr(8) |
| 17224 | .sr(1) |
| 17225 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 17226 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17227 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17228 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17229 | } |
| 17230 | } |
| 17231 | } |
| 17232 | |
| 17233 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_div_8_strided_cn) { |
| 17234 | TEST_REQUIRES_ARM_NEON; |
| 17235 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 17236 | for (size_t k = 1; k <= 80; k += 17) { |
| 17237 | GemmMicrokernelTester() |
| 17238 | .mr(1) |
| 17239 | .nr(8) |
| 17240 | .kr(8) |
| 17241 | .sr(1) |
| 17242 | .m(1) |
| 17243 | .n(n) |
| 17244 | .k(k) |
| 17245 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17246 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17247 | } |
| 17248 | } |
| 17249 | } |
| 17250 | |
| 17251 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_div_8_strided_a) { |
| 17252 | TEST_REQUIRES_ARM_NEON; |
| 17253 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 17254 | for (size_t k = 1; k <= 80; k += 17) { |
| 17255 | GemmMicrokernelTester() |
| 17256 | .mr(1) |
| 17257 | .nr(8) |
| 17258 | .kr(8) |
| 17259 | .sr(1) |
| 17260 | .m(1) |
| 17261 | .n(n) |
| 17262 | .k(k) |
| 17263 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17264 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17265 | } |
| 17266 | } |
| 17267 | } |
| 17268 | |
| 17269 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, n_div_8_subtile) { |
| 17270 | TEST_REQUIRES_ARM_NEON; |
| 17271 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 17272 | for (size_t k = 1; k <= 80; k += 17) { |
| 17273 | for (uint32_t m = 1; m <= 1; m++) { |
| 17274 | GemmMicrokernelTester() |
| 17275 | .mr(1) |
| 17276 | .nr(8) |
| 17277 | .kr(8) |
| 17278 | .sr(1) |
| 17279 | .m(m) |
| 17280 | .n(n) |
| 17281 | .k(k) |
| 17282 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17283 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17284 | } |
| 17285 | } |
| 17286 | } |
| 17287 | } |
| 17288 | |
| 17289 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, strided_cm_subtile) { |
| 17290 | TEST_REQUIRES_ARM_NEON; |
| 17291 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17292 | for (uint32_t n = 1; n <= 8; n++) { |
| 17293 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17294 | GemmMicrokernelTester() |
| 17295 | .mr(1) |
| 17296 | .nr(8) |
| 17297 | .kr(8) |
| 17298 | .sr(1) |
| 17299 | .m(m) |
| 17300 | .n(n) |
| 17301 | .k(k) |
| 17302 | .cm_stride(11) |
| 17303 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17304 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17305 | } |
| 17306 | } |
| 17307 | } |
| 17308 | } |
| 17309 | |
| 17310 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, qmin) { |
| 17311 | TEST_REQUIRES_ARM_NEON; |
| 17312 | GemmMicrokernelTester() |
| 17313 | .mr(1) |
| 17314 | .nr(8) |
| 17315 | .kr(8) |
| 17316 | .sr(1) |
| 17317 | .m(1) |
| 17318 | .n(8) |
| 17319 | .k(16) |
| 17320 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17321 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17322 | } |
| 17323 | |
| 17324 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, qmax) { |
| 17325 | TEST_REQUIRES_ARM_NEON; |
| 17326 | GemmMicrokernelTester() |
| 17327 | .mr(1) |
| 17328 | .nr(8) |
| 17329 | .kr(8) |
| 17330 | .sr(1) |
| 17331 | .m(1) |
| 17332 | .n(8) |
| 17333 | .k(16) |
| 17334 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17335 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17336 | } |
| 17337 | |
| 17338 | TEST(QS8_GEMM_MINMAX_RNDNU_1X8C8__NEON_MLAL, strided_cm) { |
| 17339 | TEST_REQUIRES_ARM_NEON; |
| 17340 | GemmMicrokernelTester() |
| 17341 | .mr(1) |
| 17342 | .nr(8) |
| 17343 | .kr(8) |
| 17344 | .sr(1) |
| 17345 | .m(1) |
| 17346 | .n(8) |
| 17347 | .k(16) |
| 17348 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17349 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x8c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17350 | } |
| 17351 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 17352 | |
| 17353 | |
| 17354 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 17355 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16) { |
| 17356 | TEST_REQUIRES_ARM_NEON; |
| 17357 | GemmMicrokernelTester() |
| 17358 | .mr(4) |
| 17359 | .nr(16) |
| 17360 | .kr(8) |
| 17361 | .sr(1) |
| 17362 | .m(4) |
| 17363 | .n(16) |
| 17364 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17365 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17366 | } |
| 17367 | |
| 17368 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, strided_cn) { |
| 17369 | TEST_REQUIRES_ARM_NEON; |
| 17370 | GemmMicrokernelTester() |
| 17371 | .mr(4) |
| 17372 | .nr(16) |
| 17373 | .kr(8) |
| 17374 | .sr(1) |
| 17375 | .m(4) |
| 17376 | .n(16) |
| 17377 | .k(16) |
| 17378 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17379 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17380 | } |
| 17381 | |
| 17382 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16_strided_a) { |
| 17383 | TEST_REQUIRES_ARM_NEON; |
| 17384 | GemmMicrokernelTester() |
| 17385 | .mr(4) |
| 17386 | .nr(16) |
| 17387 | .kr(8) |
| 17388 | .sr(1) |
| 17389 | .m(4) |
| 17390 | .n(16) |
| 17391 | .k(16) |
| 17392 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17393 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17394 | } |
| 17395 | |
| 17396 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16_subtile) { |
| 17397 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17398 | for (uint32_t n = 1; n <= 16; n++) { |
| 17399 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17400 | GemmMicrokernelTester() |
| 17401 | .mr(4) |
| 17402 | .nr(16) |
| 17403 | .kr(8) |
| 17404 | .sr(1) |
| 17405 | .m(m) |
| 17406 | .n(n) |
| 17407 | .k(16) |
| 17408 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17409 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17410 | } |
| 17411 | } |
| 17412 | } |
| 17413 | |
| 17414 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16_subtile_m) { |
| 17415 | TEST_REQUIRES_ARM_NEON; |
| 17416 | for (uint32_t m = 1; m <= 4; m++) { |
| 17417 | GemmMicrokernelTester() |
| 17418 | .mr(4) |
| 17419 | .nr(16) |
| 17420 | .kr(8) |
| 17421 | .sr(1) |
| 17422 | .m(m) |
| 17423 | .n(16) |
| 17424 | .k(16) |
| 17425 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17426 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17427 | } |
| 17428 | } |
| 17429 | |
| 17430 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_eq_16_subtile_n) { |
| 17431 | TEST_REQUIRES_ARM_NEON; |
| 17432 | for (uint32_t n = 1; n <= 16; n++) { |
| 17433 | GemmMicrokernelTester() |
| 17434 | .mr(4) |
| 17435 | .nr(16) |
| 17436 | .kr(8) |
| 17437 | .sr(1) |
| 17438 | .m(4) |
| 17439 | .n(n) |
| 17440 | .k(16) |
| 17441 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17442 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17443 | } |
| 17444 | } |
| 17445 | |
| 17446 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_lt_16) { |
| 17447 | TEST_REQUIRES_ARM_NEON; |
| 17448 | for (size_t k = 1; k < 16; k++) { |
| 17449 | GemmMicrokernelTester() |
| 17450 | .mr(4) |
| 17451 | .nr(16) |
| 17452 | .kr(8) |
| 17453 | .sr(1) |
| 17454 | .m(4) |
| 17455 | .n(16) |
| 17456 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17457 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17458 | } |
| 17459 | } |
| 17460 | |
| 17461 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_lt_16_strided_a) { |
| 17462 | TEST_REQUIRES_ARM_NEON; |
| 17463 | for (size_t k = 1; k < 16; k++) { |
| 17464 | GemmMicrokernelTester() |
| 17465 | .mr(4) |
| 17466 | .nr(16) |
| 17467 | .kr(8) |
| 17468 | .sr(1) |
| 17469 | .m(4) |
| 17470 | .n(16) |
| 17471 | .k(k) |
| 17472 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17473 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17474 | } |
| 17475 | } |
| 17476 | |
| 17477 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_lt_16_subtile) { |
| 17478 | TEST_REQUIRES_ARM_NEON; |
| 17479 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17480 | for (uint32_t n = 1; n <= 16; n++) { |
| 17481 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17482 | GemmMicrokernelTester() |
| 17483 | .mr(4) |
| 17484 | .nr(16) |
| 17485 | .kr(8) |
| 17486 | .sr(1) |
| 17487 | .m(m) |
| 17488 | .n(n) |
| 17489 | .k(k) |
| 17490 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17491 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17492 | } |
| 17493 | } |
| 17494 | } |
| 17495 | } |
| 17496 | |
| 17497 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_gt_16) { |
| 17498 | TEST_REQUIRES_ARM_NEON; |
| 17499 | for (size_t k = 17; k < 32; k++) { |
| 17500 | GemmMicrokernelTester() |
| 17501 | .mr(4) |
| 17502 | .nr(16) |
| 17503 | .kr(8) |
| 17504 | .sr(1) |
| 17505 | .m(4) |
| 17506 | .n(16) |
| 17507 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17508 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17509 | } |
| 17510 | } |
| 17511 | |
| 17512 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_gt_16_strided_a) { |
| 17513 | TEST_REQUIRES_ARM_NEON; |
| 17514 | for (size_t k = 17; k < 32; k++) { |
| 17515 | GemmMicrokernelTester() |
| 17516 | .mr(4) |
| 17517 | .nr(16) |
| 17518 | .kr(8) |
| 17519 | .sr(1) |
| 17520 | .m(4) |
| 17521 | .n(16) |
| 17522 | .k(k) |
| 17523 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17524 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17525 | } |
| 17526 | } |
| 17527 | |
| 17528 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_gt_16_subtile) { |
| 17529 | TEST_REQUIRES_ARM_NEON; |
| 17530 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17531 | for (uint32_t n = 1; n <= 16; n++) { |
| 17532 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17533 | GemmMicrokernelTester() |
| 17534 | .mr(4) |
| 17535 | .nr(16) |
| 17536 | .kr(8) |
| 17537 | .sr(1) |
| 17538 | .m(m) |
| 17539 | .n(n) |
| 17540 | .k(k) |
| 17541 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17542 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17543 | } |
| 17544 | } |
| 17545 | } |
| 17546 | } |
| 17547 | |
| 17548 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_div_16) { |
| 17549 | TEST_REQUIRES_ARM_NEON; |
| 17550 | for (size_t k = 32; k <= 160; k += 16) { |
| 17551 | GemmMicrokernelTester() |
| 17552 | .mr(4) |
| 17553 | .nr(16) |
| 17554 | .kr(8) |
| 17555 | .sr(1) |
| 17556 | .m(4) |
| 17557 | .n(16) |
| 17558 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17559 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17560 | } |
| 17561 | } |
| 17562 | |
| 17563 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_div_16_strided_a) { |
| 17564 | TEST_REQUIRES_ARM_NEON; |
| 17565 | for (size_t k = 32; k <= 160; k += 16) { |
| 17566 | GemmMicrokernelTester() |
| 17567 | .mr(4) |
| 17568 | .nr(16) |
| 17569 | .kr(8) |
| 17570 | .sr(1) |
| 17571 | .m(4) |
| 17572 | .n(16) |
| 17573 | .k(k) |
| 17574 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17575 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17576 | } |
| 17577 | } |
| 17578 | |
| 17579 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, k_div_16_subtile) { |
| 17580 | TEST_REQUIRES_ARM_NEON; |
| 17581 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17582 | for (uint32_t n = 1; n <= 16; n++) { |
| 17583 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17584 | GemmMicrokernelTester() |
| 17585 | .mr(4) |
| 17586 | .nr(16) |
| 17587 | .kr(8) |
| 17588 | .sr(1) |
| 17589 | .m(m) |
| 17590 | .n(n) |
| 17591 | .k(k) |
| 17592 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17593 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17594 | } |
| 17595 | } |
| 17596 | } |
| 17597 | } |
| 17598 | |
| 17599 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16) { |
| 17600 | TEST_REQUIRES_ARM_NEON; |
| 17601 | for (uint32_t n = 17; n < 32; n++) { |
| 17602 | for (size_t k = 1; k <= 80; k += 17) { |
| 17603 | GemmMicrokernelTester() |
| 17604 | .mr(4) |
| 17605 | .nr(16) |
| 17606 | .kr(8) |
| 17607 | .sr(1) |
| 17608 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 17609 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17610 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17611 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17612 | } |
| 17613 | } |
| 17614 | } |
| 17615 | |
| 17616 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16_strided_cn) { |
| 17617 | TEST_REQUIRES_ARM_NEON; |
| 17618 | for (uint32_t n = 17; n < 32; n++) { |
| 17619 | for (size_t k = 1; k <= 80; k += 17) { |
| 17620 | GemmMicrokernelTester() |
| 17621 | .mr(4) |
| 17622 | .nr(16) |
| 17623 | .kr(8) |
| 17624 | .sr(1) |
| 17625 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 17626 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17627 | .k(k) |
| 17628 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17629 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17630 | } |
| 17631 | } |
| 17632 | } |
| 17633 | |
| 17634 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16_strided_a) { |
| 17635 | TEST_REQUIRES_ARM_NEON; |
| 17636 | for (uint32_t n = 17; n < 32; n++) { |
| 17637 | for (size_t k = 1; k <= 80; k += 17) { |
| 17638 | GemmMicrokernelTester() |
| 17639 | .mr(4) |
| 17640 | .nr(16) |
| 17641 | .kr(8) |
| 17642 | .sr(1) |
| 17643 | .m(4) |
| 17644 | .n(n) |
| 17645 | .k(k) |
| 17646 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17647 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17648 | } |
| 17649 | } |
| 17650 | } |
| 17651 | |
| 17652 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_gt_16_subtile) { |
| 17653 | TEST_REQUIRES_ARM_NEON; |
| 17654 | for (uint32_t n = 17; n < 32; n++) { |
| 17655 | for (size_t k = 1; k <= 80; k += 17) { |
| 17656 | for (uint32_t m = 1; m <= 4; m++) { |
| 17657 | GemmMicrokernelTester() |
| 17658 | .mr(4) |
| 17659 | .nr(16) |
| 17660 | .kr(8) |
| 17661 | .sr(1) |
| 17662 | .m(m) |
| 17663 | .n(n) |
| 17664 | .k(k) |
| 17665 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17666 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17667 | } |
| 17668 | } |
| 17669 | } |
| 17670 | } |
| 17671 | |
| 17672 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16) { |
| 17673 | TEST_REQUIRES_ARM_NEON; |
| 17674 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 17675 | for (size_t k = 1; k <= 80; k += 17) { |
| 17676 | GemmMicrokernelTester() |
| 17677 | .mr(4) |
| 17678 | .nr(16) |
| 17679 | .kr(8) |
| 17680 | .sr(1) |
| 17681 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 17682 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17683 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17684 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17685 | } |
| 17686 | } |
| 17687 | } |
| 17688 | |
| 17689 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16_strided_cn) { |
| 17690 | TEST_REQUIRES_ARM_NEON; |
| 17691 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 17692 | for (size_t k = 1; k <= 80; k += 17) { |
| 17693 | GemmMicrokernelTester() |
| 17694 | .mr(4) |
| 17695 | .nr(16) |
| 17696 | .kr(8) |
| 17697 | .sr(1) |
| 17698 | .m(4) |
| 17699 | .n(n) |
| 17700 | .k(k) |
| 17701 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17702 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17703 | } |
| 17704 | } |
| 17705 | } |
| 17706 | |
| 17707 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16_strided_a) { |
| 17708 | TEST_REQUIRES_ARM_NEON; |
| 17709 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 17710 | for (size_t k = 1; k <= 80; k += 17) { |
| 17711 | GemmMicrokernelTester() |
| 17712 | .mr(4) |
| 17713 | .nr(16) |
| 17714 | .kr(8) |
| 17715 | .sr(1) |
| 17716 | .m(4) |
| 17717 | .n(n) |
| 17718 | .k(k) |
| 17719 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17720 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17721 | } |
| 17722 | } |
| 17723 | } |
| 17724 | |
| 17725 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, n_div_16_subtile) { |
| 17726 | TEST_REQUIRES_ARM_NEON; |
| 17727 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 17728 | for (size_t k = 1; k <= 80; k += 17) { |
| 17729 | for (uint32_t m = 1; m <= 4; m++) { |
| 17730 | GemmMicrokernelTester() |
| 17731 | .mr(4) |
| 17732 | .nr(16) |
| 17733 | .kr(8) |
| 17734 | .sr(1) |
| 17735 | .m(m) |
| 17736 | .n(n) |
| 17737 | .k(k) |
| 17738 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17739 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17740 | } |
| 17741 | } |
| 17742 | } |
| 17743 | } |
| 17744 | |
| 17745 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, strided_cm_subtile) { |
| 17746 | TEST_REQUIRES_ARM_NEON; |
| 17747 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17748 | for (uint32_t n = 1; n <= 16; n++) { |
| 17749 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17750 | GemmMicrokernelTester() |
| 17751 | .mr(4) |
| 17752 | .nr(16) |
| 17753 | .kr(8) |
| 17754 | .sr(1) |
| 17755 | .m(m) |
| 17756 | .n(n) |
| 17757 | .k(k) |
| 17758 | .cm_stride(19) |
| 17759 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17760 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17761 | } |
| 17762 | } |
| 17763 | } |
| 17764 | } |
| 17765 | |
| 17766 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, qmin) { |
| 17767 | TEST_REQUIRES_ARM_NEON; |
| 17768 | GemmMicrokernelTester() |
| 17769 | .mr(4) |
| 17770 | .nr(16) |
| 17771 | .kr(8) |
| 17772 | .sr(1) |
| 17773 | .m(4) |
| 17774 | .n(16) |
| 17775 | .k(16) |
| 17776 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17777 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17778 | } |
| 17779 | |
| 17780 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, qmax) { |
| 17781 | TEST_REQUIRES_ARM_NEON; |
| 17782 | GemmMicrokernelTester() |
| 17783 | .mr(4) |
| 17784 | .nr(16) |
| 17785 | .kr(8) |
| 17786 | .sr(1) |
| 17787 | .m(4) |
| 17788 | .n(16) |
| 17789 | .k(16) |
| 17790 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17791 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17792 | } |
| 17793 | |
| 17794 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C8__NEON_MLAL, strided_cm) { |
| 17795 | TEST_REQUIRES_ARM_NEON; |
| 17796 | GemmMicrokernelTester() |
| 17797 | .mr(4) |
| 17798 | .nr(16) |
| 17799 | .kr(8) |
| 17800 | .sr(1) |
| 17801 | .m(4) |
| 17802 | .n(16) |
| 17803 | .k(16) |
| 17804 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17805 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c8__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17806 | } |
| 17807 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 17808 | |
| 17809 | |
| 17810 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 17811 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16) { |
| 17812 | TEST_REQUIRES_ARM_NEON; |
| 17813 | GemmMicrokernelTester() |
| 17814 | .mr(4) |
| 17815 | .nr(16) |
| 17816 | .kr(16) |
| 17817 | .sr(1) |
| 17818 | .m(4) |
| 17819 | .n(16) |
| 17820 | .k(16) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17821 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17822 | } |
| 17823 | |
| 17824 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, strided_cn) { |
| 17825 | TEST_REQUIRES_ARM_NEON; |
| 17826 | GemmMicrokernelTester() |
| 17827 | .mr(4) |
| 17828 | .nr(16) |
| 17829 | .kr(16) |
| 17830 | .sr(1) |
| 17831 | .m(4) |
| 17832 | .n(16) |
| 17833 | .k(16) |
| 17834 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17835 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17836 | } |
| 17837 | |
| 17838 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16_strided_a) { |
| 17839 | TEST_REQUIRES_ARM_NEON; |
| 17840 | GemmMicrokernelTester() |
| 17841 | .mr(4) |
| 17842 | .nr(16) |
| 17843 | .kr(16) |
| 17844 | .sr(1) |
| 17845 | .m(4) |
| 17846 | .n(16) |
| 17847 | .k(16) |
| 17848 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17849 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17850 | } |
| 17851 | |
| 17852 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16_subtile) { |
| 17853 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17854 | for (uint32_t n = 1; n <= 16; n++) { |
| 17855 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17856 | GemmMicrokernelTester() |
| 17857 | .mr(4) |
| 17858 | .nr(16) |
| 17859 | .kr(16) |
| 17860 | .sr(1) |
| 17861 | .m(m) |
| 17862 | .n(n) |
| 17863 | .k(16) |
| 17864 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17865 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17866 | } |
| 17867 | } |
| 17868 | } |
| 17869 | |
| 17870 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16_subtile_m) { |
| 17871 | TEST_REQUIRES_ARM_NEON; |
| 17872 | for (uint32_t m = 1; m <= 4; m++) { |
| 17873 | GemmMicrokernelTester() |
| 17874 | .mr(4) |
| 17875 | .nr(16) |
| 17876 | .kr(16) |
| 17877 | .sr(1) |
| 17878 | .m(m) |
| 17879 | .n(16) |
| 17880 | .k(16) |
| 17881 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17882 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17883 | } |
| 17884 | } |
| 17885 | |
| 17886 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16_subtile_n) { |
| 17887 | TEST_REQUIRES_ARM_NEON; |
| 17888 | for (uint32_t n = 1; n <= 16; n++) { |
| 17889 | GemmMicrokernelTester() |
| 17890 | .mr(4) |
| 17891 | .nr(16) |
| 17892 | .kr(16) |
| 17893 | .sr(1) |
| 17894 | .m(4) |
| 17895 | .n(n) |
| 17896 | .k(16) |
| 17897 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17898 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17899 | } |
| 17900 | } |
| 17901 | |
| 17902 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_lt_16) { |
| 17903 | TEST_REQUIRES_ARM_NEON; |
| 17904 | for (size_t k = 1; k < 16; k++) { |
| 17905 | GemmMicrokernelTester() |
| 17906 | .mr(4) |
| 17907 | .nr(16) |
| 17908 | .kr(16) |
| 17909 | .sr(1) |
| 17910 | .m(4) |
| 17911 | .n(16) |
| 17912 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17913 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17914 | } |
| 17915 | } |
| 17916 | |
| 17917 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_lt_16_strided_a) { |
| 17918 | TEST_REQUIRES_ARM_NEON; |
| 17919 | for (size_t k = 1; k < 16; k++) { |
| 17920 | GemmMicrokernelTester() |
| 17921 | .mr(4) |
| 17922 | .nr(16) |
| 17923 | .kr(16) |
| 17924 | .sr(1) |
| 17925 | .m(4) |
| 17926 | .n(16) |
| 17927 | .k(k) |
| 17928 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17929 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17930 | } |
| 17931 | } |
| 17932 | |
| 17933 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_lt_16_subtile) { |
| 17934 | TEST_REQUIRES_ARM_NEON; |
| 17935 | for (size_t k = 1; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17936 | for (uint32_t n = 1; n <= 16; n++) { |
| 17937 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17938 | GemmMicrokernelTester() |
| 17939 | .mr(4) |
| 17940 | .nr(16) |
| 17941 | .kr(16) |
| 17942 | .sr(1) |
| 17943 | .m(m) |
| 17944 | .n(n) |
| 17945 | .k(k) |
| 17946 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17947 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17948 | } |
| 17949 | } |
| 17950 | } |
| 17951 | } |
| 17952 | |
| 17953 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_gt_16) { |
| 17954 | TEST_REQUIRES_ARM_NEON; |
| 17955 | for (size_t k = 17; k < 32; k++) { |
| 17956 | GemmMicrokernelTester() |
| 17957 | .mr(4) |
| 17958 | .nr(16) |
| 17959 | .kr(16) |
| 17960 | .sr(1) |
| 17961 | .m(4) |
| 17962 | .n(16) |
| 17963 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17964 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17965 | } |
| 17966 | } |
| 17967 | |
| 17968 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_gt_16_strided_a) { |
| 17969 | TEST_REQUIRES_ARM_NEON; |
| 17970 | for (size_t k = 17; k < 32; k++) { |
| 17971 | GemmMicrokernelTester() |
| 17972 | .mr(4) |
| 17973 | .nr(16) |
| 17974 | .kr(16) |
| 17975 | .sr(1) |
| 17976 | .m(4) |
| 17977 | .n(16) |
| 17978 | .k(k) |
| 17979 | .a_stride(37) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17980 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17981 | } |
| 17982 | } |
| 17983 | |
| 17984 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_gt_16_subtile) { |
| 17985 | TEST_REQUIRES_ARM_NEON; |
| 17986 | for (size_t k = 17; k < 32; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 17987 | for (uint32_t n = 1; n <= 16; n++) { |
| 17988 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17989 | GemmMicrokernelTester() |
| 17990 | .mr(4) |
| 17991 | .nr(16) |
| 17992 | .kr(16) |
| 17993 | .sr(1) |
| 17994 | .m(m) |
| 17995 | .n(n) |
| 17996 | .k(k) |
| 17997 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 17998 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 17999 | } |
| 18000 | } |
| 18001 | } |
| 18002 | } |
| 18003 | |
| 18004 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_div_16) { |
| 18005 | TEST_REQUIRES_ARM_NEON; |
| 18006 | for (size_t k = 32; k <= 160; k += 16) { |
| 18007 | GemmMicrokernelTester() |
| 18008 | .mr(4) |
| 18009 | .nr(16) |
| 18010 | .kr(16) |
| 18011 | .sr(1) |
| 18012 | .m(4) |
| 18013 | .n(16) |
| 18014 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18015 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18016 | } |
| 18017 | } |
| 18018 | |
| 18019 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_div_16_strided_a) { |
| 18020 | TEST_REQUIRES_ARM_NEON; |
| 18021 | for (size_t k = 32; k <= 160; k += 16) { |
| 18022 | GemmMicrokernelTester() |
| 18023 | .mr(4) |
| 18024 | .nr(16) |
| 18025 | .kr(16) |
| 18026 | .sr(1) |
| 18027 | .m(4) |
| 18028 | .n(16) |
| 18029 | .k(k) |
| 18030 | .a_stride(163) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18031 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18032 | } |
| 18033 | } |
| 18034 | |
| 18035 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_div_16_subtile) { |
| 18036 | TEST_REQUIRES_ARM_NEON; |
| 18037 | for (size_t k = 32; k <= 160; k += 16) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18038 | for (uint32_t n = 1; n <= 16; n++) { |
| 18039 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18040 | GemmMicrokernelTester() |
| 18041 | .mr(4) |
| 18042 | .nr(16) |
| 18043 | .kr(16) |
| 18044 | .sr(1) |
| 18045 | .m(m) |
| 18046 | .n(n) |
| 18047 | .k(k) |
| 18048 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18049 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18050 | } |
| 18051 | } |
| 18052 | } |
| 18053 | } |
| 18054 | |
| 18055 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16) { |
| 18056 | TEST_REQUIRES_ARM_NEON; |
| 18057 | for (uint32_t n = 17; n < 32; n++) { |
| 18058 | for (size_t k = 1; k <= 80; k += 17) { |
| 18059 | GemmMicrokernelTester() |
| 18060 | .mr(4) |
| 18061 | .nr(16) |
| 18062 | .kr(16) |
| 18063 | .sr(1) |
| 18064 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18065 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18066 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18067 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18068 | } |
| 18069 | } |
| 18070 | } |
| 18071 | |
| 18072 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16_strided_cn) { |
| 18073 | TEST_REQUIRES_ARM_NEON; |
| 18074 | for (uint32_t n = 17; n < 32; n++) { |
| 18075 | for (size_t k = 1; k <= 80; k += 17) { |
| 18076 | GemmMicrokernelTester() |
| 18077 | .mr(4) |
| 18078 | .nr(16) |
| 18079 | .kr(16) |
| 18080 | .sr(1) |
| 18081 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18082 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18083 | .k(k) |
| 18084 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18085 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18086 | } |
| 18087 | } |
| 18088 | } |
| 18089 | |
| 18090 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16_strided_a) { |
| 18091 | TEST_REQUIRES_ARM_NEON; |
| 18092 | for (uint32_t n = 17; n < 32; n++) { |
| 18093 | for (size_t k = 1; k <= 80; k += 17) { |
| 18094 | GemmMicrokernelTester() |
| 18095 | .mr(4) |
| 18096 | .nr(16) |
| 18097 | .kr(16) |
| 18098 | .sr(1) |
| 18099 | .m(4) |
| 18100 | .n(n) |
| 18101 | .k(k) |
| 18102 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18103 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18104 | } |
| 18105 | } |
| 18106 | } |
| 18107 | |
| 18108 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16_subtile) { |
| 18109 | TEST_REQUIRES_ARM_NEON; |
| 18110 | for (uint32_t n = 17; n < 32; n++) { |
| 18111 | for (size_t k = 1; k <= 80; k += 17) { |
| 18112 | for (uint32_t m = 1; m <= 4; m++) { |
| 18113 | GemmMicrokernelTester() |
| 18114 | .mr(4) |
| 18115 | .nr(16) |
| 18116 | .kr(16) |
| 18117 | .sr(1) |
| 18118 | .m(m) |
| 18119 | .n(n) |
| 18120 | .k(k) |
| 18121 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18122 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18123 | } |
| 18124 | } |
| 18125 | } |
| 18126 | } |
| 18127 | |
| 18128 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16) { |
| 18129 | TEST_REQUIRES_ARM_NEON; |
| 18130 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18131 | for (size_t k = 1; k <= 80; k += 17) { |
| 18132 | GemmMicrokernelTester() |
| 18133 | .mr(4) |
| 18134 | .nr(16) |
| 18135 | .kr(16) |
| 18136 | .sr(1) |
| 18137 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18138 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18139 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18140 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18141 | } |
| 18142 | } |
| 18143 | } |
| 18144 | |
| 18145 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16_strided_cn) { |
| 18146 | TEST_REQUIRES_ARM_NEON; |
| 18147 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18148 | for (size_t k = 1; k <= 80; k += 17) { |
| 18149 | GemmMicrokernelTester() |
| 18150 | .mr(4) |
| 18151 | .nr(16) |
| 18152 | .kr(16) |
| 18153 | .sr(1) |
| 18154 | .m(4) |
| 18155 | .n(n) |
| 18156 | .k(k) |
| 18157 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18158 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18159 | } |
| 18160 | } |
| 18161 | } |
| 18162 | |
| 18163 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16_strided_a) { |
| 18164 | TEST_REQUIRES_ARM_NEON; |
| 18165 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18166 | for (size_t k = 1; k <= 80; k += 17) { |
| 18167 | GemmMicrokernelTester() |
| 18168 | .mr(4) |
| 18169 | .nr(16) |
| 18170 | .kr(16) |
| 18171 | .sr(1) |
| 18172 | .m(4) |
| 18173 | .n(n) |
| 18174 | .k(k) |
| 18175 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18176 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18177 | } |
| 18178 | } |
| 18179 | } |
| 18180 | |
| 18181 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16_subtile) { |
| 18182 | TEST_REQUIRES_ARM_NEON; |
| 18183 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18184 | for (size_t k = 1; k <= 80; k += 17) { |
| 18185 | for (uint32_t m = 1; m <= 4; m++) { |
| 18186 | GemmMicrokernelTester() |
| 18187 | .mr(4) |
| 18188 | .nr(16) |
| 18189 | .kr(16) |
| 18190 | .sr(1) |
| 18191 | .m(m) |
| 18192 | .n(n) |
| 18193 | .k(k) |
| 18194 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18195 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18196 | } |
| 18197 | } |
| 18198 | } |
| 18199 | } |
| 18200 | |
| 18201 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, strided_cm_subtile) { |
| 18202 | TEST_REQUIRES_ARM_NEON; |
| 18203 | for (size_t k = 1; k <= 80; k += 17) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18204 | for (uint32_t n = 1; n <= 16; n++) { |
| 18205 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18206 | GemmMicrokernelTester() |
| 18207 | .mr(4) |
| 18208 | .nr(16) |
| 18209 | .kr(16) |
| 18210 | .sr(1) |
| 18211 | .m(m) |
| 18212 | .n(n) |
| 18213 | .k(k) |
| 18214 | .cm_stride(19) |
| 18215 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18216 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18217 | } |
| 18218 | } |
| 18219 | } |
| 18220 | } |
| 18221 | |
| 18222 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, qmin) { |
| 18223 | TEST_REQUIRES_ARM_NEON; |
| 18224 | GemmMicrokernelTester() |
| 18225 | .mr(4) |
| 18226 | .nr(16) |
| 18227 | .kr(16) |
| 18228 | .sr(1) |
| 18229 | .m(4) |
| 18230 | .n(16) |
| 18231 | .k(16) |
| 18232 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18233 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18234 | } |
| 18235 | |
| 18236 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, qmax) { |
| 18237 | TEST_REQUIRES_ARM_NEON; |
| 18238 | GemmMicrokernelTester() |
| 18239 | .mr(4) |
| 18240 | .nr(16) |
| 18241 | .kr(16) |
| 18242 | .sr(1) |
| 18243 | .m(4) |
| 18244 | .n(16) |
| 18245 | .k(16) |
| 18246 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18247 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18248 | } |
| 18249 | |
| 18250 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, strided_cm) { |
| 18251 | TEST_REQUIRES_ARM_NEON; |
| 18252 | GemmMicrokernelTester() |
| 18253 | .mr(4) |
| 18254 | .nr(16) |
| 18255 | .kr(16) |
| 18256 | .sr(1) |
| 18257 | .m(4) |
| 18258 | .n(16) |
| 18259 | .k(16) |
| 18260 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18261 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18262 | } |
| 18263 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 18264 | |
| 18265 | |
| 18266 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 18267 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8) { |
| 18268 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18269 | GemmMicrokernelTester() |
| 18270 | .mr(4) |
| 18271 | .nr(16) |
| 18272 | .kr(4) |
| 18273 | .sr(1) |
| 18274 | .m(4) |
| 18275 | .n(16) |
| 18276 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18277 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18278 | } |
| 18279 | |
| 18280 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, strided_cn) { |
| 18281 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18282 | GemmMicrokernelTester() |
| 18283 | .mr(4) |
| 18284 | .nr(16) |
| 18285 | .kr(4) |
| 18286 | .sr(1) |
| 18287 | .m(4) |
| 18288 | .n(16) |
| 18289 | .k(8) |
| 18290 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18291 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18292 | } |
| 18293 | |
| 18294 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8_strided_a) { |
| 18295 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18296 | GemmMicrokernelTester() |
| 18297 | .mr(4) |
| 18298 | .nr(16) |
| 18299 | .kr(4) |
| 18300 | .sr(1) |
| 18301 | .m(4) |
| 18302 | .n(16) |
| 18303 | .k(8) |
| 18304 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18305 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18306 | } |
| 18307 | |
| 18308 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8_subtile) { |
| 18309 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18310 | for (uint32_t n = 1; n <= 16; n++) { |
| 18311 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18312 | GemmMicrokernelTester() |
| 18313 | .mr(4) |
| 18314 | .nr(16) |
| 18315 | .kr(4) |
| 18316 | .sr(1) |
| 18317 | .m(m) |
| 18318 | .n(n) |
| 18319 | .k(8) |
| 18320 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18321 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18322 | } |
| 18323 | } |
| 18324 | } |
| 18325 | |
| 18326 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8_subtile_m) { |
| 18327 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18328 | for (uint32_t m = 1; m <= 4; m++) { |
| 18329 | GemmMicrokernelTester() |
| 18330 | .mr(4) |
| 18331 | .nr(16) |
| 18332 | .kr(4) |
| 18333 | .sr(1) |
| 18334 | .m(m) |
| 18335 | .n(16) |
| 18336 | .k(8) |
| 18337 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18338 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18339 | } |
| 18340 | } |
| 18341 | |
| 18342 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_eq_8_subtile_n) { |
| 18343 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18344 | for (uint32_t n = 1; n <= 16; n++) { |
| 18345 | GemmMicrokernelTester() |
| 18346 | .mr(4) |
| 18347 | .nr(16) |
| 18348 | .kr(4) |
| 18349 | .sr(1) |
| 18350 | .m(4) |
| 18351 | .n(n) |
| 18352 | .k(8) |
| 18353 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18354 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18355 | } |
| 18356 | } |
| 18357 | |
| 18358 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_lt_8) { |
| 18359 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18360 | for (size_t k = 1; k < 8; k++) { |
| 18361 | GemmMicrokernelTester() |
| 18362 | .mr(4) |
| 18363 | .nr(16) |
| 18364 | .kr(4) |
| 18365 | .sr(1) |
| 18366 | .m(4) |
| 18367 | .n(16) |
| 18368 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18369 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18370 | } |
| 18371 | } |
| 18372 | |
| 18373 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_lt_8_strided_a) { |
| 18374 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18375 | for (size_t k = 1; k < 8; k++) { |
| 18376 | GemmMicrokernelTester() |
| 18377 | .mr(4) |
| 18378 | .nr(16) |
| 18379 | .kr(4) |
| 18380 | .sr(1) |
| 18381 | .m(4) |
| 18382 | .n(16) |
| 18383 | .k(k) |
| 18384 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18385 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18386 | } |
| 18387 | } |
| 18388 | |
| 18389 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_lt_8_subtile) { |
| 18390 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18391 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18392 | for (uint32_t n = 1; n <= 16; n++) { |
| 18393 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18394 | GemmMicrokernelTester() |
| 18395 | .mr(4) |
| 18396 | .nr(16) |
| 18397 | .kr(4) |
| 18398 | .sr(1) |
| 18399 | .m(m) |
| 18400 | .n(n) |
| 18401 | .k(k) |
| 18402 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18403 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18404 | } |
| 18405 | } |
| 18406 | } |
| 18407 | } |
| 18408 | |
| 18409 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_gt_8) { |
| 18410 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18411 | for (size_t k = 9; k < 16; k++) { |
| 18412 | GemmMicrokernelTester() |
| 18413 | .mr(4) |
| 18414 | .nr(16) |
| 18415 | .kr(4) |
| 18416 | .sr(1) |
| 18417 | .m(4) |
| 18418 | .n(16) |
| 18419 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18420 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18421 | } |
| 18422 | } |
| 18423 | |
| 18424 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_gt_8_strided_a) { |
| 18425 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18426 | for (size_t k = 9; k < 16; k++) { |
| 18427 | GemmMicrokernelTester() |
| 18428 | .mr(4) |
| 18429 | .nr(16) |
| 18430 | .kr(4) |
| 18431 | .sr(1) |
| 18432 | .m(4) |
| 18433 | .n(16) |
| 18434 | .k(k) |
| 18435 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18436 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18437 | } |
| 18438 | } |
| 18439 | |
| 18440 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_gt_8_subtile) { |
| 18441 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18442 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18443 | for (uint32_t n = 1; n <= 16; n++) { |
| 18444 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18445 | GemmMicrokernelTester() |
| 18446 | .mr(4) |
| 18447 | .nr(16) |
| 18448 | .kr(4) |
| 18449 | .sr(1) |
| 18450 | .m(m) |
| 18451 | .n(n) |
| 18452 | .k(k) |
| 18453 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18454 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18455 | } |
| 18456 | } |
| 18457 | } |
| 18458 | } |
| 18459 | |
| 18460 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_div_8) { |
| 18461 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18462 | for (size_t k = 16; k <= 80; k += 8) { |
| 18463 | GemmMicrokernelTester() |
| 18464 | .mr(4) |
| 18465 | .nr(16) |
| 18466 | .kr(4) |
| 18467 | .sr(1) |
| 18468 | .m(4) |
| 18469 | .n(16) |
| 18470 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18471 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18472 | } |
| 18473 | } |
| 18474 | |
| 18475 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_div_8_strided_a) { |
| 18476 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18477 | for (size_t k = 16; k <= 80; k += 8) { |
| 18478 | GemmMicrokernelTester() |
| 18479 | .mr(4) |
| 18480 | .nr(16) |
| 18481 | .kr(4) |
| 18482 | .sr(1) |
| 18483 | .m(4) |
| 18484 | .n(16) |
| 18485 | .k(k) |
| 18486 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18487 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18488 | } |
| 18489 | } |
| 18490 | |
| 18491 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, k_div_8_subtile) { |
| 18492 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18493 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18494 | for (uint32_t n = 1; n <= 16; n++) { |
| 18495 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18496 | GemmMicrokernelTester() |
| 18497 | .mr(4) |
| 18498 | .nr(16) |
| 18499 | .kr(4) |
| 18500 | .sr(1) |
| 18501 | .m(m) |
| 18502 | .n(n) |
| 18503 | .k(k) |
| 18504 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18505 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18506 | } |
| 18507 | } |
| 18508 | } |
| 18509 | } |
| 18510 | |
| 18511 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16) { |
| 18512 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18513 | for (uint32_t n = 17; n < 32; n++) { |
| 18514 | for (size_t k = 1; k <= 40; k += 9) { |
| 18515 | GemmMicrokernelTester() |
| 18516 | .mr(4) |
| 18517 | .nr(16) |
| 18518 | .kr(4) |
| 18519 | .sr(1) |
| 18520 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18521 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18522 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18523 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18524 | } |
| 18525 | } |
| 18526 | } |
| 18527 | |
| 18528 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16_strided_cn) { |
| 18529 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18530 | for (uint32_t n = 17; n < 32; n++) { |
| 18531 | for (size_t k = 1; k <= 40; k += 9) { |
| 18532 | GemmMicrokernelTester() |
| 18533 | .mr(4) |
| 18534 | .nr(16) |
| 18535 | .kr(4) |
| 18536 | .sr(1) |
| 18537 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18538 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18539 | .k(k) |
| 18540 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18541 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18542 | } |
| 18543 | } |
| 18544 | } |
| 18545 | |
| 18546 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16_strided_a) { |
| 18547 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18548 | for (uint32_t n = 17; n < 32; n++) { |
| 18549 | for (size_t k = 1; k <= 40; k += 9) { |
| 18550 | GemmMicrokernelTester() |
| 18551 | .mr(4) |
| 18552 | .nr(16) |
| 18553 | .kr(4) |
| 18554 | .sr(1) |
| 18555 | .m(4) |
| 18556 | .n(n) |
| 18557 | .k(k) |
| 18558 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18559 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18560 | } |
| 18561 | } |
| 18562 | } |
| 18563 | |
| 18564 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_gt_16_subtile) { |
| 18565 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18566 | for (uint32_t n = 17; n < 32; n++) { |
| 18567 | for (size_t k = 1; k <= 40; k += 9) { |
| 18568 | for (uint32_t m = 1; m <= 4; m++) { |
| 18569 | GemmMicrokernelTester() |
| 18570 | .mr(4) |
| 18571 | .nr(16) |
| 18572 | .kr(4) |
| 18573 | .sr(1) |
| 18574 | .m(m) |
| 18575 | .n(n) |
| 18576 | .k(k) |
| 18577 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18578 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18579 | } |
| 18580 | } |
| 18581 | } |
| 18582 | } |
| 18583 | |
| 18584 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16) { |
| 18585 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18586 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18587 | for (size_t k = 1; k <= 40; k += 9) { |
| 18588 | GemmMicrokernelTester() |
| 18589 | .mr(4) |
| 18590 | .nr(16) |
| 18591 | .kr(4) |
| 18592 | .sr(1) |
| 18593 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18594 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18595 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18596 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18597 | } |
| 18598 | } |
| 18599 | } |
| 18600 | |
| 18601 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16_strided_cn) { |
| 18602 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18603 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18604 | for (size_t k = 1; k <= 40; k += 9) { |
| 18605 | GemmMicrokernelTester() |
| 18606 | .mr(4) |
| 18607 | .nr(16) |
| 18608 | .kr(4) |
| 18609 | .sr(1) |
| 18610 | .m(4) |
| 18611 | .n(n) |
| 18612 | .k(k) |
| 18613 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18614 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18615 | } |
| 18616 | } |
| 18617 | } |
| 18618 | |
| 18619 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16_strided_a) { |
| 18620 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18621 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18622 | for (size_t k = 1; k <= 40; k += 9) { |
| 18623 | GemmMicrokernelTester() |
| 18624 | .mr(4) |
| 18625 | .nr(16) |
| 18626 | .kr(4) |
| 18627 | .sr(1) |
| 18628 | .m(4) |
| 18629 | .n(n) |
| 18630 | .k(k) |
| 18631 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18632 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18633 | } |
| 18634 | } |
| 18635 | } |
| 18636 | |
| 18637 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, n_div_16_subtile) { |
| 18638 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18639 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 18640 | for (size_t k = 1; k <= 40; k += 9) { |
| 18641 | for (uint32_t m = 1; m <= 4; m++) { |
| 18642 | GemmMicrokernelTester() |
| 18643 | .mr(4) |
| 18644 | .nr(16) |
| 18645 | .kr(4) |
| 18646 | .sr(1) |
| 18647 | .m(m) |
| 18648 | .n(n) |
| 18649 | .k(k) |
| 18650 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18651 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18652 | } |
| 18653 | } |
| 18654 | } |
| 18655 | } |
| 18656 | |
| 18657 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, strided_cm_subtile) { |
| 18658 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18659 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18660 | for (uint32_t n = 1; n <= 16; n++) { |
| 18661 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18662 | GemmMicrokernelTester() |
| 18663 | .mr(4) |
| 18664 | .nr(16) |
| 18665 | .kr(4) |
| 18666 | .sr(1) |
| 18667 | .m(m) |
| 18668 | .n(n) |
| 18669 | .k(k) |
| 18670 | .cm_stride(19) |
| 18671 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18672 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18673 | } |
| 18674 | } |
| 18675 | } |
| 18676 | } |
| 18677 | |
| 18678 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, qmin) { |
| 18679 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18680 | GemmMicrokernelTester() |
| 18681 | .mr(4) |
| 18682 | .nr(16) |
| 18683 | .kr(4) |
| 18684 | .sr(1) |
| 18685 | .m(4) |
| 18686 | .n(16) |
| 18687 | .k(8) |
| 18688 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18689 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18690 | } |
| 18691 | |
| 18692 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, qmax) { |
| 18693 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18694 | GemmMicrokernelTester() |
| 18695 | .mr(4) |
| 18696 | .nr(16) |
| 18697 | .kr(4) |
| 18698 | .sr(1) |
| 18699 | .m(4) |
| 18700 | .n(16) |
| 18701 | .k(8) |
| 18702 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18703 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18704 | } |
| 18705 | |
| 18706 | TEST(QS8_GEMM_MINMAX_RNDNU_4X16C4__NEONDOT, strided_cm) { |
| 18707 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18708 | GemmMicrokernelTester() |
| 18709 | .mr(4) |
| 18710 | .nr(16) |
| 18711 | .kr(4) |
| 18712 | .sr(1) |
| 18713 | .m(4) |
| 18714 | .n(16) |
| 18715 | .k(8) |
| 18716 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18717 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18718 | } |
| 18719 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 18720 | |
| 18721 | |
| 18722 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 18723 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8) { |
| 18724 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18725 | GemmMicrokernelTester() |
| 18726 | .mr(6) |
| 18727 | .nr(16) |
| 18728 | .kr(4) |
| 18729 | .sr(1) |
| 18730 | .m(6) |
| 18731 | .n(16) |
| 18732 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18733 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18734 | } |
| 18735 | |
| 18736 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cn) { |
| 18737 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18738 | GemmMicrokernelTester() |
| 18739 | .mr(6) |
| 18740 | .nr(16) |
| 18741 | .kr(4) |
| 18742 | .sr(1) |
| 18743 | .m(6) |
| 18744 | .n(16) |
| 18745 | .k(8) |
| 18746 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18747 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18748 | } |
| 18749 | |
| 18750 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_strided_a) { |
| 18751 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18752 | GemmMicrokernelTester() |
| 18753 | .mr(6) |
| 18754 | .nr(16) |
| 18755 | .kr(4) |
| 18756 | .sr(1) |
| 18757 | .m(6) |
| 18758 | .n(16) |
| 18759 | .k(8) |
| 18760 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18761 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18762 | } |
| 18763 | |
| 18764 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile) { |
| 18765 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18766 | for (uint32_t n = 1; n <= 16; n++) { |
| 18767 | for (uint32_t m = 1; m <= 6; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18768 | GemmMicrokernelTester() |
| 18769 | .mr(6) |
| 18770 | .nr(16) |
| 18771 | .kr(4) |
| 18772 | .sr(1) |
| 18773 | .m(m) |
| 18774 | .n(n) |
| 18775 | .k(8) |
| 18776 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18777 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18778 | } |
| 18779 | } |
| 18780 | } |
| 18781 | |
| 18782 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile_m) { |
| 18783 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18784 | for (uint32_t m = 1; m <= 6; m++) { |
| 18785 | GemmMicrokernelTester() |
| 18786 | .mr(6) |
| 18787 | .nr(16) |
| 18788 | .kr(4) |
| 18789 | .sr(1) |
| 18790 | .m(m) |
| 18791 | .n(16) |
| 18792 | .k(8) |
| 18793 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18794 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18795 | } |
| 18796 | } |
| 18797 | |
| 18798 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile_n) { |
| 18799 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18800 | for (uint32_t n = 1; n <= 16; n++) { |
| 18801 | GemmMicrokernelTester() |
| 18802 | .mr(6) |
| 18803 | .nr(16) |
| 18804 | .kr(4) |
| 18805 | .sr(1) |
| 18806 | .m(6) |
| 18807 | .n(n) |
| 18808 | .k(8) |
| 18809 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18810 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18811 | } |
| 18812 | } |
| 18813 | |
| 18814 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8) { |
| 18815 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18816 | for (size_t k = 1; k < 8; k++) { |
| 18817 | GemmMicrokernelTester() |
| 18818 | .mr(6) |
| 18819 | .nr(16) |
| 18820 | .kr(4) |
| 18821 | .sr(1) |
| 18822 | .m(6) |
| 18823 | .n(16) |
| 18824 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18825 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18826 | } |
| 18827 | } |
| 18828 | |
| 18829 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8_strided_a) { |
| 18830 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18831 | for (size_t k = 1; k < 8; k++) { |
| 18832 | GemmMicrokernelTester() |
| 18833 | .mr(6) |
| 18834 | .nr(16) |
| 18835 | .kr(4) |
| 18836 | .sr(1) |
| 18837 | .m(6) |
| 18838 | .n(16) |
| 18839 | .k(k) |
| 18840 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18841 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18842 | } |
| 18843 | } |
| 18844 | |
| 18845 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8_subtile) { |
| 18846 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18847 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18848 | for (uint32_t n = 1; n <= 16; n++) { |
| 18849 | for (uint32_t m = 1; m <= 6; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18850 | GemmMicrokernelTester() |
| 18851 | .mr(6) |
| 18852 | .nr(16) |
| 18853 | .kr(4) |
| 18854 | .sr(1) |
| 18855 | .m(m) |
| 18856 | .n(n) |
| 18857 | .k(k) |
| 18858 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18859 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18860 | } |
| 18861 | } |
| 18862 | } |
| 18863 | } |
| 18864 | |
| 18865 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8) { |
| 18866 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18867 | for (size_t k = 9; k < 16; k++) { |
| 18868 | GemmMicrokernelTester() |
| 18869 | .mr(6) |
| 18870 | .nr(16) |
| 18871 | .kr(4) |
| 18872 | .sr(1) |
| 18873 | .m(6) |
| 18874 | .n(16) |
| 18875 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18876 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18877 | } |
| 18878 | } |
| 18879 | |
| 18880 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8_strided_a) { |
| 18881 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18882 | for (size_t k = 9; k < 16; k++) { |
| 18883 | GemmMicrokernelTester() |
| 18884 | .mr(6) |
| 18885 | .nr(16) |
| 18886 | .kr(4) |
| 18887 | .sr(1) |
| 18888 | .m(6) |
| 18889 | .n(16) |
| 18890 | .k(k) |
| 18891 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18892 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18893 | } |
| 18894 | } |
| 18895 | |
| 18896 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8_subtile) { |
| 18897 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18898 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18899 | for (uint32_t n = 1; n <= 16; n++) { |
| 18900 | for (uint32_t m = 1; m <= 6; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18901 | GemmMicrokernelTester() |
| 18902 | .mr(6) |
| 18903 | .nr(16) |
| 18904 | .kr(4) |
| 18905 | .sr(1) |
| 18906 | .m(m) |
| 18907 | .n(n) |
| 18908 | .k(k) |
| 18909 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18910 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18911 | } |
| 18912 | } |
| 18913 | } |
| 18914 | } |
| 18915 | |
| 18916 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8) { |
| 18917 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18918 | for (size_t k = 16; k <= 80; k += 8) { |
| 18919 | GemmMicrokernelTester() |
| 18920 | .mr(6) |
| 18921 | .nr(16) |
| 18922 | .kr(4) |
| 18923 | .sr(1) |
| 18924 | .m(6) |
| 18925 | .n(16) |
| 18926 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18927 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18928 | } |
| 18929 | } |
| 18930 | |
| 18931 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8_strided_a) { |
| 18932 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18933 | for (size_t k = 16; k <= 80; k += 8) { |
| 18934 | GemmMicrokernelTester() |
| 18935 | .mr(6) |
| 18936 | .nr(16) |
| 18937 | .kr(4) |
| 18938 | .sr(1) |
| 18939 | .m(6) |
| 18940 | .n(16) |
| 18941 | .k(k) |
| 18942 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18943 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18944 | } |
| 18945 | } |
| 18946 | |
| 18947 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8_subtile) { |
| 18948 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18949 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 18950 | for (uint32_t n = 1; n <= 16; n++) { |
| 18951 | for (uint32_t m = 1; m <= 6; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18952 | GemmMicrokernelTester() |
| 18953 | .mr(6) |
| 18954 | .nr(16) |
| 18955 | .kr(4) |
| 18956 | .sr(1) |
| 18957 | .m(m) |
| 18958 | .n(n) |
| 18959 | .k(k) |
| 18960 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18961 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18962 | } |
| 18963 | } |
| 18964 | } |
| 18965 | } |
| 18966 | |
| 18967 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16) { |
| 18968 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18969 | for (uint32_t n = 17; n < 32; n++) { |
| 18970 | for (size_t k = 1; k <= 40; k += 9) { |
| 18971 | GemmMicrokernelTester() |
| 18972 | .mr(6) |
| 18973 | .nr(16) |
| 18974 | .kr(4) |
| 18975 | .sr(1) |
| 18976 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18977 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18978 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18979 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18980 | } |
| 18981 | } |
| 18982 | } |
| 18983 | |
| 18984 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_strided_cn) { |
| 18985 | TEST_REQUIRES_ARM_NEON_DOT; |
| 18986 | for (uint32_t n = 17; n < 32; n++) { |
| 18987 | for (size_t k = 1; k <= 40; k += 9) { |
| 18988 | GemmMicrokernelTester() |
| 18989 | .mr(6) |
| 18990 | .nr(16) |
| 18991 | .kr(4) |
| 18992 | .sr(1) |
| 18993 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 18994 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18995 | .k(k) |
| 18996 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 18997 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 18998 | } |
| 18999 | } |
| 19000 | } |
| 19001 | |
| 19002 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_strided_a) { |
| 19003 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19004 | for (uint32_t n = 17; n < 32; n++) { |
| 19005 | for (size_t k = 1; k <= 40; k += 9) { |
| 19006 | GemmMicrokernelTester() |
| 19007 | .mr(6) |
| 19008 | .nr(16) |
| 19009 | .kr(4) |
| 19010 | .sr(1) |
| 19011 | .m(6) |
| 19012 | .n(n) |
| 19013 | .k(k) |
| 19014 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19015 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19016 | } |
| 19017 | } |
| 19018 | } |
| 19019 | |
| 19020 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_subtile) { |
| 19021 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19022 | for (uint32_t n = 17; n < 32; n++) { |
| 19023 | for (size_t k = 1; k <= 40; k += 9) { |
| 19024 | for (uint32_t m = 1; m <= 6; m++) { |
| 19025 | GemmMicrokernelTester() |
| 19026 | .mr(6) |
| 19027 | .nr(16) |
| 19028 | .kr(4) |
| 19029 | .sr(1) |
| 19030 | .m(m) |
| 19031 | .n(n) |
| 19032 | .k(k) |
| 19033 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19034 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19035 | } |
| 19036 | } |
| 19037 | } |
| 19038 | } |
| 19039 | |
| 19040 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16) { |
| 19041 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19042 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19043 | for (size_t k = 1; k <= 40; k += 9) { |
| 19044 | GemmMicrokernelTester() |
| 19045 | .mr(6) |
| 19046 | .nr(16) |
| 19047 | .kr(4) |
| 19048 | .sr(1) |
| 19049 | .m(6) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 19050 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19051 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19052 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19053 | } |
| 19054 | } |
| 19055 | } |
| 19056 | |
| 19057 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_strided_cn) { |
| 19058 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19059 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19060 | for (size_t k = 1; k <= 40; k += 9) { |
| 19061 | GemmMicrokernelTester() |
| 19062 | .mr(6) |
| 19063 | .nr(16) |
| 19064 | .kr(4) |
| 19065 | .sr(1) |
| 19066 | .m(6) |
| 19067 | .n(n) |
| 19068 | .k(k) |
| 19069 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19070 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19071 | } |
| 19072 | } |
| 19073 | } |
| 19074 | |
| 19075 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_strided_a) { |
| 19076 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19077 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19078 | for (size_t k = 1; k <= 40; k += 9) { |
| 19079 | GemmMicrokernelTester() |
| 19080 | .mr(6) |
| 19081 | .nr(16) |
| 19082 | .kr(4) |
| 19083 | .sr(1) |
| 19084 | .m(6) |
| 19085 | .n(n) |
| 19086 | .k(k) |
| 19087 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19088 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19089 | } |
| 19090 | } |
| 19091 | } |
| 19092 | |
| 19093 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_subtile) { |
| 19094 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19095 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19096 | for (size_t k = 1; k <= 40; k += 9) { |
| 19097 | for (uint32_t m = 1; m <= 6; m++) { |
| 19098 | GemmMicrokernelTester() |
| 19099 | .mr(6) |
| 19100 | .nr(16) |
| 19101 | .kr(4) |
| 19102 | .sr(1) |
| 19103 | .m(m) |
| 19104 | .n(n) |
| 19105 | .k(k) |
| 19106 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19107 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19108 | } |
| 19109 | } |
| 19110 | } |
| 19111 | } |
| 19112 | |
| 19113 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cm_subtile) { |
| 19114 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19115 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19116 | for (uint32_t n = 1; n <= 16; n++) { |
| 19117 | for (uint32_t m = 1; m <= 6; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19118 | GemmMicrokernelTester() |
| 19119 | .mr(6) |
| 19120 | .nr(16) |
| 19121 | .kr(4) |
| 19122 | .sr(1) |
| 19123 | .m(m) |
| 19124 | .n(n) |
| 19125 | .k(k) |
| 19126 | .cm_stride(19) |
| 19127 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19128 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19129 | } |
| 19130 | } |
| 19131 | } |
| 19132 | } |
| 19133 | |
| 19134 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, qmin) { |
| 19135 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19136 | GemmMicrokernelTester() |
| 19137 | .mr(6) |
| 19138 | .nr(16) |
| 19139 | .kr(4) |
| 19140 | .sr(1) |
| 19141 | .m(6) |
| 19142 | .n(16) |
| 19143 | .k(8) |
| 19144 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19145 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19146 | } |
| 19147 | |
| 19148 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, qmax) { |
| 19149 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19150 | GemmMicrokernelTester() |
| 19151 | .mr(6) |
| 19152 | .nr(16) |
| 19153 | .kr(4) |
| 19154 | .sr(1) |
| 19155 | .m(6) |
| 19156 | .n(16) |
| 19157 | .k(8) |
| 19158 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19159 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19160 | } |
| 19161 | |
| 19162 | TEST(QS8_GEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cm) { |
| 19163 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19164 | GemmMicrokernelTester() |
| 19165 | .mr(6) |
| 19166 | .nr(16) |
| 19167 | .kr(4) |
| 19168 | .sr(1) |
| 19169 | .m(6) |
| 19170 | .n(16) |
| 19171 | .k(8) |
| 19172 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19173 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19174 | } |
| 19175 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 19176 | |
| 19177 | |
| 19178 | #if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 19179 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_eq_8) { |
| 19180 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19181 | GemmMicrokernelTester() |
| 19182 | .mr(8) |
| 19183 | .nr(16) |
| 19184 | .kr(4) |
| 19185 | .sr(1) |
| 19186 | .m(8) |
| 19187 | .n(16) |
| 19188 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19189 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19190 | } |
| 19191 | |
| 19192 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, strided_cn) { |
| 19193 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19194 | GemmMicrokernelTester() |
| 19195 | .mr(8) |
| 19196 | .nr(16) |
| 19197 | .kr(4) |
| 19198 | .sr(1) |
| 19199 | .m(8) |
| 19200 | .n(16) |
| 19201 | .k(8) |
| 19202 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19203 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19204 | } |
| 19205 | |
| 19206 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_eq_8_strided_a) { |
| 19207 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19208 | GemmMicrokernelTester() |
| 19209 | .mr(8) |
| 19210 | .nr(16) |
| 19211 | .kr(4) |
| 19212 | .sr(1) |
| 19213 | .m(8) |
| 19214 | .n(16) |
| 19215 | .k(8) |
| 19216 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19217 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19218 | } |
| 19219 | |
| 19220 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_eq_8_subtile) { |
| 19221 | TEST_REQUIRES_ARM_NEON_DOT; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19222 | for (uint32_t n = 1; n <= 16; n++) { |
| 19223 | for (uint32_t m = 1; m <= 8; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19224 | GemmMicrokernelTester() |
| 19225 | .mr(8) |
| 19226 | .nr(16) |
| 19227 | .kr(4) |
| 19228 | .sr(1) |
| 19229 | .m(m) |
| 19230 | .n(n) |
| 19231 | .k(8) |
| 19232 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19233 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19234 | } |
| 19235 | } |
| 19236 | } |
| 19237 | |
| 19238 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_eq_8_subtile_m) { |
| 19239 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19240 | for (uint32_t m = 1; m <= 8; m++) { |
| 19241 | GemmMicrokernelTester() |
| 19242 | .mr(8) |
| 19243 | .nr(16) |
| 19244 | .kr(4) |
| 19245 | .sr(1) |
| 19246 | .m(m) |
| 19247 | .n(16) |
| 19248 | .k(8) |
| 19249 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19250 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19251 | } |
| 19252 | } |
| 19253 | |
| 19254 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_eq_8_subtile_n) { |
| 19255 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19256 | for (uint32_t n = 1; n <= 16; n++) { |
| 19257 | GemmMicrokernelTester() |
| 19258 | .mr(8) |
| 19259 | .nr(16) |
| 19260 | .kr(4) |
| 19261 | .sr(1) |
| 19262 | .m(8) |
| 19263 | .n(n) |
| 19264 | .k(8) |
| 19265 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19266 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19267 | } |
| 19268 | } |
| 19269 | |
| 19270 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_lt_8) { |
| 19271 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19272 | for (size_t k = 1; k < 8; k++) { |
| 19273 | GemmMicrokernelTester() |
| 19274 | .mr(8) |
| 19275 | .nr(16) |
| 19276 | .kr(4) |
| 19277 | .sr(1) |
| 19278 | .m(8) |
| 19279 | .n(16) |
| 19280 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19281 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19282 | } |
| 19283 | } |
| 19284 | |
| 19285 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_lt_8_strided_a) { |
| 19286 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19287 | for (size_t k = 1; k < 8; k++) { |
| 19288 | GemmMicrokernelTester() |
| 19289 | .mr(8) |
| 19290 | .nr(16) |
| 19291 | .kr(4) |
| 19292 | .sr(1) |
| 19293 | .m(8) |
| 19294 | .n(16) |
| 19295 | .k(k) |
| 19296 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19297 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19298 | } |
| 19299 | } |
| 19300 | |
| 19301 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_lt_8_subtile) { |
| 19302 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19303 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19304 | for (uint32_t n = 1; n <= 16; n++) { |
| 19305 | for (uint32_t m = 1; m <= 8; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19306 | GemmMicrokernelTester() |
| 19307 | .mr(8) |
| 19308 | .nr(16) |
| 19309 | .kr(4) |
| 19310 | .sr(1) |
| 19311 | .m(m) |
| 19312 | .n(n) |
| 19313 | .k(k) |
| 19314 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19315 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19316 | } |
| 19317 | } |
| 19318 | } |
| 19319 | } |
| 19320 | |
| 19321 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_gt_8) { |
| 19322 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19323 | for (size_t k = 9; k < 16; k++) { |
| 19324 | GemmMicrokernelTester() |
| 19325 | .mr(8) |
| 19326 | .nr(16) |
| 19327 | .kr(4) |
| 19328 | .sr(1) |
| 19329 | .m(8) |
| 19330 | .n(16) |
| 19331 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19332 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19333 | } |
| 19334 | } |
| 19335 | |
| 19336 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_gt_8_strided_a) { |
| 19337 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19338 | for (size_t k = 9; k < 16; k++) { |
| 19339 | GemmMicrokernelTester() |
| 19340 | .mr(8) |
| 19341 | .nr(16) |
| 19342 | .kr(4) |
| 19343 | .sr(1) |
| 19344 | .m(8) |
| 19345 | .n(16) |
| 19346 | .k(k) |
| 19347 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19348 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19349 | } |
| 19350 | } |
| 19351 | |
| 19352 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_gt_8_subtile) { |
| 19353 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19354 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19355 | for (uint32_t n = 1; n <= 16; n++) { |
| 19356 | for (uint32_t m = 1; m <= 8; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19357 | GemmMicrokernelTester() |
| 19358 | .mr(8) |
| 19359 | .nr(16) |
| 19360 | .kr(4) |
| 19361 | .sr(1) |
| 19362 | .m(m) |
| 19363 | .n(n) |
| 19364 | .k(k) |
| 19365 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19366 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19367 | } |
| 19368 | } |
| 19369 | } |
| 19370 | } |
| 19371 | |
| 19372 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_div_8) { |
| 19373 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19374 | for (size_t k = 16; k <= 80; k += 8) { |
| 19375 | GemmMicrokernelTester() |
| 19376 | .mr(8) |
| 19377 | .nr(16) |
| 19378 | .kr(4) |
| 19379 | .sr(1) |
| 19380 | .m(8) |
| 19381 | .n(16) |
| 19382 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19383 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19384 | } |
| 19385 | } |
| 19386 | |
| 19387 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_div_8_strided_a) { |
| 19388 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19389 | for (size_t k = 16; k <= 80; k += 8) { |
| 19390 | GemmMicrokernelTester() |
| 19391 | .mr(8) |
| 19392 | .nr(16) |
| 19393 | .kr(4) |
| 19394 | .sr(1) |
| 19395 | .m(8) |
| 19396 | .n(16) |
| 19397 | .k(k) |
| 19398 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19399 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19400 | } |
| 19401 | } |
| 19402 | |
| 19403 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, k_div_8_subtile) { |
| 19404 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19405 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19406 | for (uint32_t n = 1; n <= 16; n++) { |
| 19407 | for (uint32_t m = 1; m <= 8; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19408 | GemmMicrokernelTester() |
| 19409 | .mr(8) |
| 19410 | .nr(16) |
| 19411 | .kr(4) |
| 19412 | .sr(1) |
| 19413 | .m(m) |
| 19414 | .n(n) |
| 19415 | .k(k) |
| 19416 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19417 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19418 | } |
| 19419 | } |
| 19420 | } |
| 19421 | } |
| 19422 | |
| 19423 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_gt_16) { |
| 19424 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19425 | for (uint32_t n = 17; n < 32; n++) { |
| 19426 | for (size_t k = 1; k <= 40; k += 9) { |
| 19427 | GemmMicrokernelTester() |
| 19428 | .mr(8) |
| 19429 | .nr(16) |
| 19430 | .kr(4) |
| 19431 | .sr(1) |
| 19432 | .m(8) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 19433 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19434 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19435 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19436 | } |
| 19437 | } |
| 19438 | } |
| 19439 | |
| 19440 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_gt_16_strided_cn) { |
| 19441 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19442 | for (uint32_t n = 17; n < 32; n++) { |
| 19443 | for (size_t k = 1; k <= 40; k += 9) { |
| 19444 | GemmMicrokernelTester() |
| 19445 | .mr(8) |
| 19446 | .nr(16) |
| 19447 | .kr(4) |
| 19448 | .sr(1) |
| 19449 | .m(8) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 19450 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19451 | .k(k) |
| 19452 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19453 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19454 | } |
| 19455 | } |
| 19456 | } |
| 19457 | |
| 19458 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_gt_16_strided_a) { |
| 19459 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19460 | for (uint32_t n = 17; n < 32; n++) { |
| 19461 | for (size_t k = 1; k <= 40; k += 9) { |
| 19462 | GemmMicrokernelTester() |
| 19463 | .mr(8) |
| 19464 | .nr(16) |
| 19465 | .kr(4) |
| 19466 | .sr(1) |
| 19467 | .m(8) |
| 19468 | .n(n) |
| 19469 | .k(k) |
| 19470 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19471 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19472 | } |
| 19473 | } |
| 19474 | } |
| 19475 | |
| 19476 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_gt_16_subtile) { |
| 19477 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19478 | for (uint32_t n = 17; n < 32; n++) { |
| 19479 | for (size_t k = 1; k <= 40; k += 9) { |
| 19480 | for (uint32_t m = 1; m <= 8; m++) { |
| 19481 | GemmMicrokernelTester() |
| 19482 | .mr(8) |
| 19483 | .nr(16) |
| 19484 | .kr(4) |
| 19485 | .sr(1) |
| 19486 | .m(m) |
| 19487 | .n(n) |
| 19488 | .k(k) |
| 19489 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19490 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19491 | } |
| 19492 | } |
| 19493 | } |
| 19494 | } |
| 19495 | |
| 19496 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_div_16) { |
| 19497 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19498 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19499 | for (size_t k = 1; k <= 40; k += 9) { |
| 19500 | GemmMicrokernelTester() |
| 19501 | .mr(8) |
| 19502 | .nr(16) |
| 19503 | .kr(4) |
| 19504 | .sr(1) |
| 19505 | .m(8) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 19506 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19507 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19508 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19509 | } |
| 19510 | } |
| 19511 | } |
| 19512 | |
| 19513 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_div_16_strided_cn) { |
| 19514 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19515 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19516 | for (size_t k = 1; k <= 40; k += 9) { |
| 19517 | GemmMicrokernelTester() |
| 19518 | .mr(8) |
| 19519 | .nr(16) |
| 19520 | .kr(4) |
| 19521 | .sr(1) |
| 19522 | .m(8) |
| 19523 | .n(n) |
| 19524 | .k(k) |
| 19525 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19526 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19527 | } |
| 19528 | } |
| 19529 | } |
| 19530 | |
| 19531 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_div_16_strided_a) { |
| 19532 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19533 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19534 | for (size_t k = 1; k <= 40; k += 9) { |
| 19535 | GemmMicrokernelTester() |
| 19536 | .mr(8) |
| 19537 | .nr(16) |
| 19538 | .kr(4) |
| 19539 | .sr(1) |
| 19540 | .m(8) |
| 19541 | .n(n) |
| 19542 | .k(k) |
| 19543 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19544 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19545 | } |
| 19546 | } |
| 19547 | } |
| 19548 | |
| 19549 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, n_div_16_subtile) { |
| 19550 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19551 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19552 | for (size_t k = 1; k <= 40; k += 9) { |
| 19553 | for (uint32_t m = 1; m <= 8; m++) { |
| 19554 | GemmMicrokernelTester() |
| 19555 | .mr(8) |
| 19556 | .nr(16) |
| 19557 | .kr(4) |
| 19558 | .sr(1) |
| 19559 | .m(m) |
| 19560 | .n(n) |
| 19561 | .k(k) |
| 19562 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19563 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19564 | } |
| 19565 | } |
| 19566 | } |
| 19567 | } |
| 19568 | |
| 19569 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, strided_cm_subtile) { |
| 19570 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19571 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19572 | for (uint32_t n = 1; n <= 16; n++) { |
| 19573 | for (uint32_t m = 1; m <= 8; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19574 | GemmMicrokernelTester() |
| 19575 | .mr(8) |
| 19576 | .nr(16) |
| 19577 | .kr(4) |
| 19578 | .sr(1) |
| 19579 | .m(m) |
| 19580 | .n(n) |
| 19581 | .k(k) |
| 19582 | .cm_stride(19) |
| 19583 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19584 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19585 | } |
| 19586 | } |
| 19587 | } |
| 19588 | } |
| 19589 | |
| 19590 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, qmin) { |
| 19591 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19592 | GemmMicrokernelTester() |
| 19593 | .mr(8) |
| 19594 | .nr(16) |
| 19595 | .kr(4) |
| 19596 | .sr(1) |
| 19597 | .m(8) |
| 19598 | .n(16) |
| 19599 | .k(8) |
| 19600 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19601 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19602 | } |
| 19603 | |
| 19604 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, qmax) { |
| 19605 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19606 | GemmMicrokernelTester() |
| 19607 | .mr(8) |
| 19608 | .nr(16) |
| 19609 | .kr(4) |
| 19610 | .sr(1) |
| 19611 | .m(8) |
| 19612 | .n(16) |
| 19613 | .k(8) |
| 19614 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19615 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19616 | } |
| 19617 | |
| 19618 | TEST(QS8_GEMM_MINMAX_RNDNU_8X16C4__NEONDOT, strided_cm) { |
| 19619 | TEST_REQUIRES_ARM_NEON_DOT; |
| 19620 | GemmMicrokernelTester() |
| 19621 | .mr(8) |
| 19622 | .nr(16) |
| 19623 | .kr(4) |
| 19624 | .sr(1) |
| 19625 | .m(8) |
| 19626 | .n(16) |
| 19627 | .k(8) |
| 19628 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19629 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_8x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19630 | } |
| 19631 | #endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64 |
| 19632 | |
| 19633 | |
| 19634 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 19635 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8) { |
| 19636 | TEST_REQUIRES_ARM_NEON; |
| 19637 | GemmMicrokernelTester() |
| 19638 | .mr(1) |
| 19639 | .nr(16) |
| 19640 | .kr(1) |
| 19641 | .sr(1) |
| 19642 | .m(1) |
| 19643 | .n(16) |
| 19644 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19645 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19646 | } |
| 19647 | |
| 19648 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cn) { |
| 19649 | TEST_REQUIRES_ARM_NEON; |
| 19650 | GemmMicrokernelTester() |
| 19651 | .mr(1) |
| 19652 | .nr(16) |
| 19653 | .kr(1) |
| 19654 | .sr(1) |
| 19655 | .m(1) |
| 19656 | .n(16) |
| 19657 | .k(8) |
| 19658 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19659 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19660 | } |
| 19661 | |
| 19662 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_strided_a) { |
| 19663 | TEST_REQUIRES_ARM_NEON; |
| 19664 | GemmMicrokernelTester() |
| 19665 | .mr(1) |
| 19666 | .nr(16) |
| 19667 | .kr(1) |
| 19668 | .sr(1) |
| 19669 | .m(1) |
| 19670 | .n(16) |
| 19671 | .k(8) |
| 19672 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19673 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19674 | } |
| 19675 | |
| 19676 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile) { |
| 19677 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19678 | for (uint32_t n = 1; n <= 16; n++) { |
| 19679 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19680 | GemmMicrokernelTester() |
| 19681 | .mr(1) |
| 19682 | .nr(16) |
| 19683 | .kr(1) |
| 19684 | .sr(1) |
| 19685 | .m(m) |
| 19686 | .n(n) |
| 19687 | .k(8) |
| 19688 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19689 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19690 | } |
| 19691 | } |
| 19692 | } |
| 19693 | |
| 19694 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile_m) { |
| 19695 | TEST_REQUIRES_ARM_NEON; |
| 19696 | for (uint32_t m = 1; m <= 1; m++) { |
| 19697 | GemmMicrokernelTester() |
| 19698 | .mr(1) |
| 19699 | .nr(16) |
| 19700 | .kr(1) |
| 19701 | .sr(1) |
| 19702 | .m(m) |
| 19703 | .n(16) |
| 19704 | .k(8) |
| 19705 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19706 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19707 | } |
| 19708 | } |
| 19709 | |
| 19710 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_eq_8_subtile_n) { |
| 19711 | TEST_REQUIRES_ARM_NEON; |
| 19712 | for (uint32_t n = 1; n <= 16; n++) { |
| 19713 | GemmMicrokernelTester() |
| 19714 | .mr(1) |
| 19715 | .nr(16) |
| 19716 | .kr(1) |
| 19717 | .sr(1) |
| 19718 | .m(1) |
| 19719 | .n(n) |
| 19720 | .k(8) |
| 19721 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19722 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19723 | } |
| 19724 | } |
| 19725 | |
| 19726 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8) { |
| 19727 | TEST_REQUIRES_ARM_NEON; |
| 19728 | for (size_t k = 1; k < 8; k++) { |
| 19729 | GemmMicrokernelTester() |
| 19730 | .mr(1) |
| 19731 | .nr(16) |
| 19732 | .kr(1) |
| 19733 | .sr(1) |
| 19734 | .m(1) |
| 19735 | .n(16) |
| 19736 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19737 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19738 | } |
| 19739 | } |
| 19740 | |
| 19741 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8_strided_a) { |
| 19742 | TEST_REQUIRES_ARM_NEON; |
| 19743 | for (size_t k = 1; k < 8; k++) { |
| 19744 | GemmMicrokernelTester() |
| 19745 | .mr(1) |
| 19746 | .nr(16) |
| 19747 | .kr(1) |
| 19748 | .sr(1) |
| 19749 | .m(1) |
| 19750 | .n(16) |
| 19751 | .k(k) |
| 19752 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19753 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19754 | } |
| 19755 | } |
| 19756 | |
| 19757 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_lt_8_subtile) { |
| 19758 | TEST_REQUIRES_ARM_NEON; |
| 19759 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19760 | for (uint32_t n = 1; n <= 16; n++) { |
| 19761 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19762 | GemmMicrokernelTester() |
| 19763 | .mr(1) |
| 19764 | .nr(16) |
| 19765 | .kr(1) |
| 19766 | .sr(1) |
| 19767 | .m(m) |
| 19768 | .n(n) |
| 19769 | .k(k) |
| 19770 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19771 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19772 | } |
| 19773 | } |
| 19774 | } |
| 19775 | } |
| 19776 | |
| 19777 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8) { |
| 19778 | TEST_REQUIRES_ARM_NEON; |
| 19779 | for (size_t k = 9; k < 16; k++) { |
| 19780 | GemmMicrokernelTester() |
| 19781 | .mr(1) |
| 19782 | .nr(16) |
| 19783 | .kr(1) |
| 19784 | .sr(1) |
| 19785 | .m(1) |
| 19786 | .n(16) |
| 19787 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19788 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19789 | } |
| 19790 | } |
| 19791 | |
| 19792 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8_strided_a) { |
| 19793 | TEST_REQUIRES_ARM_NEON; |
| 19794 | for (size_t k = 9; k < 16; k++) { |
| 19795 | GemmMicrokernelTester() |
| 19796 | .mr(1) |
| 19797 | .nr(16) |
| 19798 | .kr(1) |
| 19799 | .sr(1) |
| 19800 | .m(1) |
| 19801 | .n(16) |
| 19802 | .k(k) |
| 19803 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19804 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19805 | } |
| 19806 | } |
| 19807 | |
| 19808 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_gt_8_subtile) { |
| 19809 | TEST_REQUIRES_ARM_NEON; |
| 19810 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19811 | for (uint32_t n = 1; n <= 16; n++) { |
| 19812 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19813 | GemmMicrokernelTester() |
| 19814 | .mr(1) |
| 19815 | .nr(16) |
| 19816 | .kr(1) |
| 19817 | .sr(1) |
| 19818 | .m(m) |
| 19819 | .n(n) |
| 19820 | .k(k) |
| 19821 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19822 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19823 | } |
| 19824 | } |
| 19825 | } |
| 19826 | } |
| 19827 | |
| 19828 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8) { |
| 19829 | TEST_REQUIRES_ARM_NEON; |
| 19830 | for (size_t k = 16; k <= 80; k += 8) { |
| 19831 | GemmMicrokernelTester() |
| 19832 | .mr(1) |
| 19833 | .nr(16) |
| 19834 | .kr(1) |
| 19835 | .sr(1) |
| 19836 | .m(1) |
| 19837 | .n(16) |
| 19838 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19839 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19840 | } |
| 19841 | } |
| 19842 | |
| 19843 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8_strided_a) { |
| 19844 | TEST_REQUIRES_ARM_NEON; |
| 19845 | for (size_t k = 16; k <= 80; k += 8) { |
| 19846 | GemmMicrokernelTester() |
| 19847 | .mr(1) |
| 19848 | .nr(16) |
| 19849 | .kr(1) |
| 19850 | .sr(1) |
| 19851 | .m(1) |
| 19852 | .n(16) |
| 19853 | .k(k) |
| 19854 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19855 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19856 | } |
| 19857 | } |
| 19858 | |
| 19859 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, k_div_8_subtile) { |
| 19860 | TEST_REQUIRES_ARM_NEON; |
| 19861 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 19862 | for (uint32_t n = 1; n <= 16; n++) { |
| 19863 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19864 | GemmMicrokernelTester() |
| 19865 | .mr(1) |
| 19866 | .nr(16) |
| 19867 | .kr(1) |
| 19868 | .sr(1) |
| 19869 | .m(m) |
| 19870 | .n(n) |
| 19871 | .k(k) |
| 19872 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19873 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19874 | } |
| 19875 | } |
| 19876 | } |
| 19877 | } |
| 19878 | |
| 19879 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16) { |
| 19880 | TEST_REQUIRES_ARM_NEON; |
| 19881 | for (uint32_t n = 17; n < 32; n++) { |
| 19882 | for (size_t k = 1; k <= 40; k += 9) { |
| 19883 | GemmMicrokernelTester() |
| 19884 | .mr(1) |
| 19885 | .nr(16) |
| 19886 | .kr(1) |
| 19887 | .sr(1) |
| 19888 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 19889 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19890 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19891 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19892 | } |
| 19893 | } |
| 19894 | } |
| 19895 | |
| 19896 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_strided_cn) { |
| 19897 | TEST_REQUIRES_ARM_NEON; |
| 19898 | for (uint32_t n = 17; n < 32; n++) { |
| 19899 | for (size_t k = 1; k <= 40; k += 9) { |
| 19900 | GemmMicrokernelTester() |
| 19901 | .mr(1) |
| 19902 | .nr(16) |
| 19903 | .kr(1) |
| 19904 | .sr(1) |
| 19905 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 19906 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19907 | .k(k) |
| 19908 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19909 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19910 | } |
| 19911 | } |
| 19912 | } |
| 19913 | |
| 19914 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_strided_a) { |
| 19915 | TEST_REQUIRES_ARM_NEON; |
| 19916 | for (uint32_t n = 17; n < 32; n++) { |
| 19917 | for (size_t k = 1; k <= 40; k += 9) { |
| 19918 | GemmMicrokernelTester() |
| 19919 | .mr(1) |
| 19920 | .nr(16) |
| 19921 | .kr(1) |
| 19922 | .sr(1) |
| 19923 | .m(1) |
| 19924 | .n(n) |
| 19925 | .k(k) |
| 19926 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19927 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19928 | } |
| 19929 | } |
| 19930 | } |
| 19931 | |
| 19932 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_gt_16_subtile) { |
| 19933 | TEST_REQUIRES_ARM_NEON; |
| 19934 | for (uint32_t n = 17; n < 32; n++) { |
| 19935 | for (size_t k = 1; k <= 40; k += 9) { |
| 19936 | for (uint32_t m = 1; m <= 1; m++) { |
| 19937 | GemmMicrokernelTester() |
| 19938 | .mr(1) |
| 19939 | .nr(16) |
| 19940 | .kr(1) |
| 19941 | .sr(1) |
| 19942 | .m(m) |
| 19943 | .n(n) |
| 19944 | .k(k) |
| 19945 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19946 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19947 | } |
| 19948 | } |
| 19949 | } |
| 19950 | } |
| 19951 | |
| 19952 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16) { |
| 19953 | TEST_REQUIRES_ARM_NEON; |
| 19954 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19955 | for (size_t k = 1; k <= 40; k += 9) { |
| 19956 | GemmMicrokernelTester() |
| 19957 | .mr(1) |
| 19958 | .nr(16) |
| 19959 | .kr(1) |
| 19960 | .sr(1) |
| 19961 | .m(1) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 19962 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19963 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19964 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19965 | } |
| 19966 | } |
| 19967 | } |
| 19968 | |
| 19969 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_strided_cn) { |
| 19970 | TEST_REQUIRES_ARM_NEON; |
| 19971 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19972 | for (size_t k = 1; k <= 40; k += 9) { |
| 19973 | GemmMicrokernelTester() |
| 19974 | .mr(1) |
| 19975 | .nr(16) |
| 19976 | .kr(1) |
| 19977 | .sr(1) |
| 19978 | .m(1) |
| 19979 | .n(n) |
| 19980 | .k(k) |
| 19981 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 19982 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 19983 | } |
| 19984 | } |
| 19985 | } |
| 19986 | |
| 19987 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_strided_a) { |
| 19988 | TEST_REQUIRES_ARM_NEON; |
| 19989 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 19990 | for (size_t k = 1; k <= 40; k += 9) { |
| 19991 | GemmMicrokernelTester() |
| 19992 | .mr(1) |
| 19993 | .nr(16) |
| 19994 | .kr(1) |
| 19995 | .sr(1) |
| 19996 | .m(1) |
| 19997 | .n(n) |
| 19998 | .k(k) |
| 19999 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20000 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20001 | } |
| 20002 | } |
| 20003 | } |
| 20004 | |
| 20005 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, n_div_16_subtile) { |
| 20006 | TEST_REQUIRES_ARM_NEON; |
| 20007 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 20008 | for (size_t k = 1; k <= 40; k += 9) { |
| 20009 | for (uint32_t m = 1; m <= 1; m++) { |
| 20010 | GemmMicrokernelTester() |
| 20011 | .mr(1) |
| 20012 | .nr(16) |
| 20013 | .kr(1) |
| 20014 | .sr(1) |
| 20015 | .m(m) |
| 20016 | .n(n) |
| 20017 | .k(k) |
| 20018 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20019 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20020 | } |
| 20021 | } |
| 20022 | } |
| 20023 | } |
| 20024 | |
| 20025 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cm_subtile) { |
| 20026 | TEST_REQUIRES_ARM_NEON; |
| 20027 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20028 | for (uint32_t n = 1; n <= 16; n++) { |
| 20029 | for (uint32_t m = 1; m <= 1; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20030 | GemmMicrokernelTester() |
| 20031 | .mr(1) |
| 20032 | .nr(16) |
| 20033 | .kr(1) |
| 20034 | .sr(1) |
| 20035 | .m(m) |
| 20036 | .n(n) |
| 20037 | .k(k) |
| 20038 | .cm_stride(19) |
| 20039 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20040 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20041 | } |
| 20042 | } |
| 20043 | } |
| 20044 | } |
| 20045 | |
| 20046 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, qmin) { |
| 20047 | TEST_REQUIRES_ARM_NEON; |
| 20048 | GemmMicrokernelTester() |
| 20049 | .mr(1) |
| 20050 | .nr(16) |
| 20051 | .kr(1) |
| 20052 | .sr(1) |
| 20053 | .m(1) |
| 20054 | .n(16) |
| 20055 | .k(8) |
| 20056 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20057 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20058 | } |
| 20059 | |
| 20060 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, qmax) { |
| 20061 | TEST_REQUIRES_ARM_NEON; |
| 20062 | GemmMicrokernelTester() |
| 20063 | .mr(1) |
| 20064 | .nr(16) |
| 20065 | .kr(1) |
| 20066 | .sr(1) |
| 20067 | .m(1) |
| 20068 | .n(16) |
| 20069 | .k(8) |
| 20070 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20071 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20072 | } |
| 20073 | |
| 20074 | TEST(QS8_GEMM_MINMAX_RNDNU_1X16__NEON_MLAL_LANE, strided_cm) { |
| 20075 | TEST_REQUIRES_ARM_NEON; |
| 20076 | GemmMicrokernelTester() |
| 20077 | .mr(1) |
| 20078 | .nr(16) |
| 20079 | .kr(1) |
| 20080 | .sr(1) |
| 20081 | .m(1) |
| 20082 | .n(16) |
| 20083 | .k(8) |
| 20084 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20085 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_1x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20086 | } |
| 20087 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 20088 | |
| 20089 | |
| 20090 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 20091 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_eq_8) { |
| 20092 | TEST_REQUIRES_ARM_NEON; |
| 20093 | GemmMicrokernelTester() |
| 20094 | .mr(2) |
| 20095 | .nr(8) |
| 20096 | .kr(1) |
| 20097 | .sr(1) |
| 20098 | .m(2) |
| 20099 | .n(8) |
| 20100 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20101 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20102 | } |
| 20103 | |
| 20104 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, strided_cn) { |
| 20105 | TEST_REQUIRES_ARM_NEON; |
| 20106 | GemmMicrokernelTester() |
| 20107 | .mr(2) |
| 20108 | .nr(8) |
| 20109 | .kr(1) |
| 20110 | .sr(1) |
| 20111 | .m(2) |
| 20112 | .n(8) |
| 20113 | .k(8) |
| 20114 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20115 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20116 | } |
| 20117 | |
| 20118 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_eq_8_strided_a) { |
| 20119 | TEST_REQUIRES_ARM_NEON; |
| 20120 | GemmMicrokernelTester() |
| 20121 | .mr(2) |
| 20122 | .nr(8) |
| 20123 | .kr(1) |
| 20124 | .sr(1) |
| 20125 | .m(2) |
| 20126 | .n(8) |
| 20127 | .k(8) |
| 20128 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20129 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20130 | } |
| 20131 | |
| 20132 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile) { |
| 20133 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20134 | for (uint32_t n = 1; n <= 8; n++) { |
| 20135 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20136 | GemmMicrokernelTester() |
| 20137 | .mr(2) |
| 20138 | .nr(8) |
| 20139 | .kr(1) |
| 20140 | .sr(1) |
| 20141 | .m(m) |
| 20142 | .n(n) |
| 20143 | .k(8) |
| 20144 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20145 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20146 | } |
| 20147 | } |
| 20148 | } |
| 20149 | |
| 20150 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_m) { |
| 20151 | TEST_REQUIRES_ARM_NEON; |
| 20152 | for (uint32_t m = 1; m <= 2; m++) { |
| 20153 | GemmMicrokernelTester() |
| 20154 | .mr(2) |
| 20155 | .nr(8) |
| 20156 | .kr(1) |
| 20157 | .sr(1) |
| 20158 | .m(m) |
| 20159 | .n(8) |
| 20160 | .k(8) |
| 20161 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20162 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20163 | } |
| 20164 | } |
| 20165 | |
| 20166 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_n) { |
| 20167 | TEST_REQUIRES_ARM_NEON; |
| 20168 | for (uint32_t n = 1; n <= 8; n++) { |
| 20169 | GemmMicrokernelTester() |
| 20170 | .mr(2) |
| 20171 | .nr(8) |
| 20172 | .kr(1) |
| 20173 | .sr(1) |
| 20174 | .m(2) |
| 20175 | .n(n) |
| 20176 | .k(8) |
| 20177 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20178 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20179 | } |
| 20180 | } |
| 20181 | |
| 20182 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_lt_8) { |
| 20183 | TEST_REQUIRES_ARM_NEON; |
| 20184 | for (size_t k = 1; k < 8; k++) { |
| 20185 | GemmMicrokernelTester() |
| 20186 | .mr(2) |
| 20187 | .nr(8) |
| 20188 | .kr(1) |
| 20189 | .sr(1) |
| 20190 | .m(2) |
| 20191 | .n(8) |
| 20192 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20193 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20194 | } |
| 20195 | } |
| 20196 | |
| 20197 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_lt_8_strided_a) { |
| 20198 | TEST_REQUIRES_ARM_NEON; |
| 20199 | for (size_t k = 1; k < 8; k++) { |
| 20200 | GemmMicrokernelTester() |
| 20201 | .mr(2) |
| 20202 | .nr(8) |
| 20203 | .kr(1) |
| 20204 | .sr(1) |
| 20205 | .m(2) |
| 20206 | .n(8) |
| 20207 | .k(k) |
| 20208 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20209 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20210 | } |
| 20211 | } |
| 20212 | |
| 20213 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_lt_8_subtile) { |
| 20214 | TEST_REQUIRES_ARM_NEON; |
| 20215 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20216 | for (uint32_t n = 1; n <= 8; n++) { |
| 20217 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20218 | GemmMicrokernelTester() |
| 20219 | .mr(2) |
| 20220 | .nr(8) |
| 20221 | .kr(1) |
| 20222 | .sr(1) |
| 20223 | .m(m) |
| 20224 | .n(n) |
| 20225 | .k(k) |
| 20226 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20227 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20228 | } |
| 20229 | } |
| 20230 | } |
| 20231 | } |
| 20232 | |
| 20233 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_gt_8) { |
| 20234 | TEST_REQUIRES_ARM_NEON; |
| 20235 | for (size_t k = 9; k < 16; k++) { |
| 20236 | GemmMicrokernelTester() |
| 20237 | .mr(2) |
| 20238 | .nr(8) |
| 20239 | .kr(1) |
| 20240 | .sr(1) |
| 20241 | .m(2) |
| 20242 | .n(8) |
| 20243 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20244 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20245 | } |
| 20246 | } |
| 20247 | |
| 20248 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_gt_8_strided_a) { |
| 20249 | TEST_REQUIRES_ARM_NEON; |
| 20250 | for (size_t k = 9; k < 16; k++) { |
| 20251 | GemmMicrokernelTester() |
| 20252 | .mr(2) |
| 20253 | .nr(8) |
| 20254 | .kr(1) |
| 20255 | .sr(1) |
| 20256 | .m(2) |
| 20257 | .n(8) |
| 20258 | .k(k) |
| 20259 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20260 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20261 | } |
| 20262 | } |
| 20263 | |
| 20264 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_gt_8_subtile) { |
| 20265 | TEST_REQUIRES_ARM_NEON; |
| 20266 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20267 | for (uint32_t n = 1; n <= 8; n++) { |
| 20268 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20269 | GemmMicrokernelTester() |
| 20270 | .mr(2) |
| 20271 | .nr(8) |
| 20272 | .kr(1) |
| 20273 | .sr(1) |
| 20274 | .m(m) |
| 20275 | .n(n) |
| 20276 | .k(k) |
| 20277 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20278 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20279 | } |
| 20280 | } |
| 20281 | } |
| 20282 | } |
| 20283 | |
| 20284 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_div_8) { |
| 20285 | TEST_REQUIRES_ARM_NEON; |
| 20286 | for (size_t k = 16; k <= 80; k += 8) { |
| 20287 | GemmMicrokernelTester() |
| 20288 | .mr(2) |
| 20289 | .nr(8) |
| 20290 | .kr(1) |
| 20291 | .sr(1) |
| 20292 | .m(2) |
| 20293 | .n(8) |
| 20294 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20295 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20296 | } |
| 20297 | } |
| 20298 | |
| 20299 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_div_8_strided_a) { |
| 20300 | TEST_REQUIRES_ARM_NEON; |
| 20301 | for (size_t k = 16; k <= 80; k += 8) { |
| 20302 | GemmMicrokernelTester() |
| 20303 | .mr(2) |
| 20304 | .nr(8) |
| 20305 | .kr(1) |
| 20306 | .sr(1) |
| 20307 | .m(2) |
| 20308 | .n(8) |
| 20309 | .k(k) |
| 20310 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20311 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20312 | } |
| 20313 | } |
| 20314 | |
| 20315 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, k_div_8_subtile) { |
| 20316 | TEST_REQUIRES_ARM_NEON; |
| 20317 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20318 | for (uint32_t n = 1; n <= 8; n++) { |
| 20319 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20320 | GemmMicrokernelTester() |
| 20321 | .mr(2) |
| 20322 | .nr(8) |
| 20323 | .kr(1) |
| 20324 | .sr(1) |
| 20325 | .m(m) |
| 20326 | .n(n) |
| 20327 | .k(k) |
| 20328 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20329 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20330 | } |
| 20331 | } |
| 20332 | } |
| 20333 | } |
| 20334 | |
| 20335 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_gt_8) { |
| 20336 | TEST_REQUIRES_ARM_NEON; |
| 20337 | for (uint32_t n = 9; n < 16; n++) { |
| 20338 | for (size_t k = 1; k <= 40; k += 9) { |
| 20339 | GemmMicrokernelTester() |
| 20340 | .mr(2) |
| 20341 | .nr(8) |
| 20342 | .kr(1) |
| 20343 | .sr(1) |
| 20344 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 20345 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20346 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20347 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20348 | } |
| 20349 | } |
| 20350 | } |
| 20351 | |
| 20352 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_gt_8_strided_cn) { |
| 20353 | TEST_REQUIRES_ARM_NEON; |
| 20354 | for (uint32_t n = 9; n < 16; n++) { |
| 20355 | for (size_t k = 1; k <= 40; k += 9) { |
| 20356 | GemmMicrokernelTester() |
| 20357 | .mr(2) |
| 20358 | .nr(8) |
| 20359 | .kr(1) |
| 20360 | .sr(1) |
| 20361 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 20362 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20363 | .k(k) |
| 20364 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20365 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20366 | } |
| 20367 | } |
| 20368 | } |
| 20369 | |
| 20370 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_gt_8_strided_a) { |
| 20371 | TEST_REQUIRES_ARM_NEON; |
| 20372 | for (uint32_t n = 9; n < 16; n++) { |
| 20373 | for (size_t k = 1; k <= 40; k += 9) { |
| 20374 | GemmMicrokernelTester() |
| 20375 | .mr(2) |
| 20376 | .nr(8) |
| 20377 | .kr(1) |
| 20378 | .sr(1) |
| 20379 | .m(2) |
| 20380 | .n(n) |
| 20381 | .k(k) |
| 20382 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20383 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20384 | } |
| 20385 | } |
| 20386 | } |
| 20387 | |
| 20388 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_gt_8_subtile) { |
| 20389 | TEST_REQUIRES_ARM_NEON; |
| 20390 | for (uint32_t n = 9; n < 16; n++) { |
| 20391 | for (size_t k = 1; k <= 40; k += 9) { |
| 20392 | for (uint32_t m = 1; m <= 2; m++) { |
| 20393 | GemmMicrokernelTester() |
| 20394 | .mr(2) |
| 20395 | .nr(8) |
| 20396 | .kr(1) |
| 20397 | .sr(1) |
| 20398 | .m(m) |
| 20399 | .n(n) |
| 20400 | .k(k) |
| 20401 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20402 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20403 | } |
| 20404 | } |
| 20405 | } |
| 20406 | } |
| 20407 | |
| 20408 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_div_8) { |
| 20409 | TEST_REQUIRES_ARM_NEON; |
| 20410 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20411 | for (size_t k = 1; k <= 40; k += 9) { |
| 20412 | GemmMicrokernelTester() |
| 20413 | .mr(2) |
| 20414 | .nr(8) |
| 20415 | .kr(1) |
| 20416 | .sr(1) |
| 20417 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 20418 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20419 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20420 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20421 | } |
| 20422 | } |
| 20423 | } |
| 20424 | |
| 20425 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_div_8_strided_cn) { |
| 20426 | TEST_REQUIRES_ARM_NEON; |
| 20427 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20428 | for (size_t k = 1; k <= 40; k += 9) { |
| 20429 | GemmMicrokernelTester() |
| 20430 | .mr(2) |
| 20431 | .nr(8) |
| 20432 | .kr(1) |
| 20433 | .sr(1) |
| 20434 | .m(2) |
| 20435 | .n(n) |
| 20436 | .k(k) |
| 20437 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20438 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20439 | } |
| 20440 | } |
| 20441 | } |
| 20442 | |
| 20443 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_div_8_strided_a) { |
| 20444 | TEST_REQUIRES_ARM_NEON; |
| 20445 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20446 | for (size_t k = 1; k <= 40; k += 9) { |
| 20447 | GemmMicrokernelTester() |
| 20448 | .mr(2) |
| 20449 | .nr(8) |
| 20450 | .kr(1) |
| 20451 | .sr(1) |
| 20452 | .m(2) |
| 20453 | .n(n) |
| 20454 | .k(k) |
| 20455 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20456 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20457 | } |
| 20458 | } |
| 20459 | } |
| 20460 | |
| 20461 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, n_div_8_subtile) { |
| 20462 | TEST_REQUIRES_ARM_NEON; |
| 20463 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20464 | for (size_t k = 1; k <= 40; k += 9) { |
| 20465 | for (uint32_t m = 1; m <= 2; m++) { |
| 20466 | GemmMicrokernelTester() |
| 20467 | .mr(2) |
| 20468 | .nr(8) |
| 20469 | .kr(1) |
| 20470 | .sr(1) |
| 20471 | .m(m) |
| 20472 | .n(n) |
| 20473 | .k(k) |
| 20474 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20475 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20476 | } |
| 20477 | } |
| 20478 | } |
| 20479 | } |
| 20480 | |
| 20481 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, strided_cm_subtile) { |
| 20482 | TEST_REQUIRES_ARM_NEON; |
| 20483 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20484 | for (uint32_t n = 1; n <= 8; n++) { |
| 20485 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20486 | GemmMicrokernelTester() |
| 20487 | .mr(2) |
| 20488 | .nr(8) |
| 20489 | .kr(1) |
| 20490 | .sr(1) |
| 20491 | .m(m) |
| 20492 | .n(n) |
| 20493 | .k(k) |
| 20494 | .cm_stride(11) |
| 20495 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20496 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20497 | } |
| 20498 | } |
| 20499 | } |
| 20500 | } |
| 20501 | |
| 20502 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, qmin) { |
| 20503 | TEST_REQUIRES_ARM_NEON; |
| 20504 | GemmMicrokernelTester() |
| 20505 | .mr(2) |
| 20506 | .nr(8) |
| 20507 | .kr(1) |
| 20508 | .sr(1) |
| 20509 | .m(2) |
| 20510 | .n(8) |
| 20511 | .k(8) |
| 20512 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20513 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20514 | } |
| 20515 | |
| 20516 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, qmax) { |
| 20517 | TEST_REQUIRES_ARM_NEON; |
| 20518 | GemmMicrokernelTester() |
| 20519 | .mr(2) |
| 20520 | .nr(8) |
| 20521 | .kr(1) |
| 20522 | .sr(1) |
| 20523 | .m(2) |
| 20524 | .n(8) |
| 20525 | .k(8) |
| 20526 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20527 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20528 | } |
| 20529 | |
| 20530 | TEST(QS8_GEMM_MINMAX_RNDNU_2X8__NEON_MULL_ADDW_DUP, strided_cm) { |
| 20531 | TEST_REQUIRES_ARM_NEON; |
| 20532 | GemmMicrokernelTester() |
| 20533 | .mr(2) |
| 20534 | .nr(8) |
| 20535 | .kr(1) |
| 20536 | .sr(1) |
| 20537 | .m(2) |
| 20538 | .n(8) |
| 20539 | .k(8) |
| 20540 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20541 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20542 | } |
| 20543 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 20544 | |
| 20545 | |
| 20546 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 20547 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8) { |
| 20548 | TEST_REQUIRES_ARM_NEON; |
| 20549 | GemmMicrokernelTester() |
| 20550 | .mr(4) |
| 20551 | .nr(8) |
| 20552 | .kr(1) |
| 20553 | .sr(1) |
| 20554 | .m(4) |
| 20555 | .n(8) |
| 20556 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20557 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20558 | } |
| 20559 | |
| 20560 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, strided_cn) { |
| 20561 | TEST_REQUIRES_ARM_NEON; |
| 20562 | GemmMicrokernelTester() |
| 20563 | .mr(4) |
| 20564 | .nr(8) |
| 20565 | .kr(1) |
| 20566 | .sr(1) |
| 20567 | .m(4) |
| 20568 | .n(8) |
| 20569 | .k(8) |
| 20570 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20571 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20572 | } |
| 20573 | |
| 20574 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8_strided_a) { |
| 20575 | TEST_REQUIRES_ARM_NEON; |
| 20576 | GemmMicrokernelTester() |
| 20577 | .mr(4) |
| 20578 | .nr(8) |
| 20579 | .kr(1) |
| 20580 | .sr(1) |
| 20581 | .m(4) |
| 20582 | .n(8) |
| 20583 | .k(8) |
| 20584 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20585 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20586 | } |
| 20587 | |
| 20588 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile) { |
| 20589 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20590 | for (uint32_t n = 1; n <= 8; n++) { |
| 20591 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20592 | GemmMicrokernelTester() |
| 20593 | .mr(4) |
| 20594 | .nr(8) |
| 20595 | .kr(1) |
| 20596 | .sr(1) |
| 20597 | .m(m) |
| 20598 | .n(n) |
| 20599 | .k(8) |
| 20600 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20601 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20602 | } |
| 20603 | } |
| 20604 | } |
| 20605 | |
| 20606 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_m) { |
| 20607 | TEST_REQUIRES_ARM_NEON; |
| 20608 | for (uint32_t m = 1; m <= 4; m++) { |
| 20609 | GemmMicrokernelTester() |
| 20610 | .mr(4) |
| 20611 | .nr(8) |
| 20612 | .kr(1) |
| 20613 | .sr(1) |
| 20614 | .m(m) |
| 20615 | .n(8) |
| 20616 | .k(8) |
| 20617 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20618 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20619 | } |
| 20620 | } |
| 20621 | |
| 20622 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_eq_8_subtile_n) { |
| 20623 | TEST_REQUIRES_ARM_NEON; |
| 20624 | for (uint32_t n = 1; n <= 8; n++) { |
| 20625 | GemmMicrokernelTester() |
| 20626 | .mr(4) |
| 20627 | .nr(8) |
| 20628 | .kr(1) |
| 20629 | .sr(1) |
| 20630 | .m(4) |
| 20631 | .n(n) |
| 20632 | .k(8) |
| 20633 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20634 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20635 | } |
| 20636 | } |
| 20637 | |
| 20638 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_lt_8) { |
| 20639 | TEST_REQUIRES_ARM_NEON; |
| 20640 | for (size_t k = 1; k < 8; k++) { |
| 20641 | GemmMicrokernelTester() |
| 20642 | .mr(4) |
| 20643 | .nr(8) |
| 20644 | .kr(1) |
| 20645 | .sr(1) |
| 20646 | .m(4) |
| 20647 | .n(8) |
| 20648 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20649 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20650 | } |
| 20651 | } |
| 20652 | |
| 20653 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_lt_8_strided_a) { |
| 20654 | TEST_REQUIRES_ARM_NEON; |
| 20655 | for (size_t k = 1; k < 8; k++) { |
| 20656 | GemmMicrokernelTester() |
| 20657 | .mr(4) |
| 20658 | .nr(8) |
| 20659 | .kr(1) |
| 20660 | .sr(1) |
| 20661 | .m(4) |
| 20662 | .n(8) |
| 20663 | .k(k) |
| 20664 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20665 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20666 | } |
| 20667 | } |
| 20668 | |
| 20669 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_lt_8_subtile) { |
| 20670 | TEST_REQUIRES_ARM_NEON; |
| 20671 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20672 | for (uint32_t n = 1; n <= 8; n++) { |
| 20673 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20674 | GemmMicrokernelTester() |
| 20675 | .mr(4) |
| 20676 | .nr(8) |
| 20677 | .kr(1) |
| 20678 | .sr(1) |
| 20679 | .m(m) |
| 20680 | .n(n) |
| 20681 | .k(k) |
| 20682 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20683 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20684 | } |
| 20685 | } |
| 20686 | } |
| 20687 | } |
| 20688 | |
| 20689 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_gt_8) { |
| 20690 | TEST_REQUIRES_ARM_NEON; |
| 20691 | for (size_t k = 9; k < 16; k++) { |
| 20692 | GemmMicrokernelTester() |
| 20693 | .mr(4) |
| 20694 | .nr(8) |
| 20695 | .kr(1) |
| 20696 | .sr(1) |
| 20697 | .m(4) |
| 20698 | .n(8) |
| 20699 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20700 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20701 | } |
| 20702 | } |
| 20703 | |
| 20704 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_gt_8_strided_a) { |
| 20705 | TEST_REQUIRES_ARM_NEON; |
| 20706 | for (size_t k = 9; k < 16; k++) { |
| 20707 | GemmMicrokernelTester() |
| 20708 | .mr(4) |
| 20709 | .nr(8) |
| 20710 | .kr(1) |
| 20711 | .sr(1) |
| 20712 | .m(4) |
| 20713 | .n(8) |
| 20714 | .k(k) |
| 20715 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20716 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20717 | } |
| 20718 | } |
| 20719 | |
| 20720 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_gt_8_subtile) { |
| 20721 | TEST_REQUIRES_ARM_NEON; |
| 20722 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20723 | for (uint32_t n = 1; n <= 8; n++) { |
| 20724 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20725 | GemmMicrokernelTester() |
| 20726 | .mr(4) |
| 20727 | .nr(8) |
| 20728 | .kr(1) |
| 20729 | .sr(1) |
| 20730 | .m(m) |
| 20731 | .n(n) |
| 20732 | .k(k) |
| 20733 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20734 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20735 | } |
| 20736 | } |
| 20737 | } |
| 20738 | } |
| 20739 | |
| 20740 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_div_8) { |
| 20741 | TEST_REQUIRES_ARM_NEON; |
| 20742 | for (size_t k = 16; k <= 80; k += 8) { |
| 20743 | GemmMicrokernelTester() |
| 20744 | .mr(4) |
| 20745 | .nr(8) |
| 20746 | .kr(1) |
| 20747 | .sr(1) |
| 20748 | .m(4) |
| 20749 | .n(8) |
| 20750 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20751 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20752 | } |
| 20753 | } |
| 20754 | |
| 20755 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_div_8_strided_a) { |
| 20756 | TEST_REQUIRES_ARM_NEON; |
| 20757 | for (size_t k = 16; k <= 80; k += 8) { |
| 20758 | GemmMicrokernelTester() |
| 20759 | .mr(4) |
| 20760 | .nr(8) |
| 20761 | .kr(1) |
| 20762 | .sr(1) |
| 20763 | .m(4) |
| 20764 | .n(8) |
| 20765 | .k(k) |
| 20766 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20767 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20768 | } |
| 20769 | } |
| 20770 | |
| 20771 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, k_div_8_subtile) { |
| 20772 | TEST_REQUIRES_ARM_NEON; |
| 20773 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20774 | for (uint32_t n = 1; n <= 8; n++) { |
| 20775 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20776 | GemmMicrokernelTester() |
| 20777 | .mr(4) |
| 20778 | .nr(8) |
| 20779 | .kr(1) |
| 20780 | .sr(1) |
| 20781 | .m(m) |
| 20782 | .n(n) |
| 20783 | .k(k) |
| 20784 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20785 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20786 | } |
| 20787 | } |
| 20788 | } |
| 20789 | } |
| 20790 | |
| 20791 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8) { |
| 20792 | TEST_REQUIRES_ARM_NEON; |
| 20793 | for (uint32_t n = 9; n < 16; n++) { |
| 20794 | for (size_t k = 1; k <= 40; k += 9) { |
| 20795 | GemmMicrokernelTester() |
| 20796 | .mr(4) |
| 20797 | .nr(8) |
| 20798 | .kr(1) |
| 20799 | .sr(1) |
| 20800 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 20801 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20802 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20803 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20804 | } |
| 20805 | } |
| 20806 | } |
| 20807 | |
| 20808 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8_strided_cn) { |
| 20809 | TEST_REQUIRES_ARM_NEON; |
| 20810 | for (uint32_t n = 9; n < 16; n++) { |
| 20811 | for (size_t k = 1; k <= 40; k += 9) { |
| 20812 | GemmMicrokernelTester() |
| 20813 | .mr(4) |
| 20814 | .nr(8) |
| 20815 | .kr(1) |
| 20816 | .sr(1) |
| 20817 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 20818 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20819 | .k(k) |
| 20820 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20821 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20822 | } |
| 20823 | } |
| 20824 | } |
| 20825 | |
| 20826 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8_strided_a) { |
| 20827 | TEST_REQUIRES_ARM_NEON; |
| 20828 | for (uint32_t n = 9; n < 16; n++) { |
| 20829 | for (size_t k = 1; k <= 40; k += 9) { |
| 20830 | GemmMicrokernelTester() |
| 20831 | .mr(4) |
| 20832 | .nr(8) |
| 20833 | .kr(1) |
| 20834 | .sr(1) |
| 20835 | .m(4) |
| 20836 | .n(n) |
| 20837 | .k(k) |
| 20838 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20839 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20840 | } |
| 20841 | } |
| 20842 | } |
| 20843 | |
| 20844 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_gt_8_subtile) { |
| 20845 | TEST_REQUIRES_ARM_NEON; |
| 20846 | for (uint32_t n = 9; n < 16; n++) { |
| 20847 | for (size_t k = 1; k <= 40; k += 9) { |
| 20848 | for (uint32_t m = 1; m <= 4; m++) { |
| 20849 | GemmMicrokernelTester() |
| 20850 | .mr(4) |
| 20851 | .nr(8) |
| 20852 | .kr(1) |
| 20853 | .sr(1) |
| 20854 | .m(m) |
| 20855 | .n(n) |
| 20856 | .k(k) |
| 20857 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20858 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20859 | } |
| 20860 | } |
| 20861 | } |
| 20862 | } |
| 20863 | |
| 20864 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8) { |
| 20865 | TEST_REQUIRES_ARM_NEON; |
| 20866 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20867 | for (size_t k = 1; k <= 40; k += 9) { |
| 20868 | GemmMicrokernelTester() |
| 20869 | .mr(4) |
| 20870 | .nr(8) |
| 20871 | .kr(1) |
| 20872 | .sr(1) |
| 20873 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 20874 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20875 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20876 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20877 | } |
| 20878 | } |
| 20879 | } |
| 20880 | |
| 20881 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8_strided_cn) { |
| 20882 | TEST_REQUIRES_ARM_NEON; |
| 20883 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20884 | for (size_t k = 1; k <= 40; k += 9) { |
| 20885 | GemmMicrokernelTester() |
| 20886 | .mr(4) |
| 20887 | .nr(8) |
| 20888 | .kr(1) |
| 20889 | .sr(1) |
| 20890 | .m(4) |
| 20891 | .n(n) |
| 20892 | .k(k) |
| 20893 | .cn_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20894 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20895 | } |
| 20896 | } |
| 20897 | } |
| 20898 | |
| 20899 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8_strided_a) { |
| 20900 | TEST_REQUIRES_ARM_NEON; |
| 20901 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20902 | for (size_t k = 1; k <= 40; k += 9) { |
| 20903 | GemmMicrokernelTester() |
| 20904 | .mr(4) |
| 20905 | .nr(8) |
| 20906 | .kr(1) |
| 20907 | .sr(1) |
| 20908 | .m(4) |
| 20909 | .n(n) |
| 20910 | .k(k) |
| 20911 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20912 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20913 | } |
| 20914 | } |
| 20915 | } |
| 20916 | |
| 20917 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, n_div_8_subtile) { |
| 20918 | TEST_REQUIRES_ARM_NEON; |
| 20919 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 20920 | for (size_t k = 1; k <= 40; k += 9) { |
| 20921 | for (uint32_t m = 1; m <= 4; m++) { |
| 20922 | GemmMicrokernelTester() |
| 20923 | .mr(4) |
| 20924 | .nr(8) |
| 20925 | .kr(1) |
| 20926 | .sr(1) |
| 20927 | .m(m) |
| 20928 | .n(n) |
| 20929 | .k(k) |
| 20930 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20931 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20932 | } |
| 20933 | } |
| 20934 | } |
| 20935 | } |
| 20936 | |
| 20937 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, strided_cm_subtile) { |
| 20938 | TEST_REQUIRES_ARM_NEON; |
| 20939 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 20940 | for (uint32_t n = 1; n <= 8; n++) { |
| 20941 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20942 | GemmMicrokernelTester() |
| 20943 | .mr(4) |
| 20944 | .nr(8) |
| 20945 | .kr(1) |
| 20946 | .sr(1) |
| 20947 | .m(m) |
| 20948 | .n(n) |
| 20949 | .k(k) |
| 20950 | .cm_stride(11) |
| 20951 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20952 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20953 | } |
| 20954 | } |
| 20955 | } |
| 20956 | } |
| 20957 | |
| 20958 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, qmin) { |
| 20959 | TEST_REQUIRES_ARM_NEON; |
| 20960 | GemmMicrokernelTester() |
| 20961 | .mr(4) |
| 20962 | .nr(8) |
| 20963 | .kr(1) |
| 20964 | .sr(1) |
| 20965 | .m(4) |
| 20966 | .n(8) |
| 20967 | .k(8) |
| 20968 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20969 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20970 | } |
| 20971 | |
| 20972 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, qmax) { |
| 20973 | TEST_REQUIRES_ARM_NEON; |
| 20974 | GemmMicrokernelTester() |
| 20975 | .mr(4) |
| 20976 | .nr(8) |
| 20977 | .kr(1) |
| 20978 | .sr(1) |
| 20979 | .m(4) |
| 20980 | .n(8) |
| 20981 | .k(8) |
| 20982 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20983 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20984 | } |
| 20985 | |
| 20986 | TEST(QS8_GEMM_MINMAX_RNDNU_4X8__NEON_MULL_ADDW_DUP, strided_cm) { |
| 20987 | TEST_REQUIRES_ARM_NEON; |
| 20988 | GemmMicrokernelTester() |
| 20989 | .mr(4) |
| 20990 | .nr(8) |
| 20991 | .kr(1) |
| 20992 | .sr(1) |
| 20993 | .m(4) |
| 20994 | .n(8) |
| 20995 | .k(8) |
| 20996 | .cm_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 20997 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_4x8__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 20998 | } |
| 20999 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 21000 | |
| 21001 | |
| 21002 | #if XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 21003 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_eq_8) { |
| 21004 | TEST_REQUIRES_ARM_NEON; |
| 21005 | GemmMicrokernelTester() |
| 21006 | .mr(2) |
| 21007 | .nr(16) |
| 21008 | .kr(1) |
| 21009 | .sr(1) |
| 21010 | .m(2) |
| 21011 | .n(16) |
| 21012 | .k(8) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21013 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21014 | } |
| 21015 | |
| 21016 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, strided_cn) { |
| 21017 | TEST_REQUIRES_ARM_NEON; |
| 21018 | GemmMicrokernelTester() |
| 21019 | .mr(2) |
| 21020 | .nr(16) |
| 21021 | .kr(1) |
| 21022 | .sr(1) |
| 21023 | .m(2) |
| 21024 | .n(16) |
| 21025 | .k(8) |
| 21026 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21027 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21028 | } |
| 21029 | |
| 21030 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_eq_8_strided_a) { |
| 21031 | TEST_REQUIRES_ARM_NEON; |
| 21032 | GemmMicrokernelTester() |
| 21033 | .mr(2) |
| 21034 | .nr(16) |
| 21035 | .kr(1) |
| 21036 | .sr(1) |
| 21037 | .m(2) |
| 21038 | .n(16) |
| 21039 | .k(8) |
| 21040 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21041 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21042 | } |
| 21043 | |
| 21044 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_eq_8_subtile) { |
| 21045 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21046 | for (uint32_t n = 1; n <= 16; n++) { |
| 21047 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21048 | GemmMicrokernelTester() |
| 21049 | .mr(2) |
| 21050 | .nr(16) |
| 21051 | .kr(1) |
| 21052 | .sr(1) |
| 21053 | .m(m) |
| 21054 | .n(n) |
| 21055 | .k(8) |
| 21056 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21057 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21058 | } |
| 21059 | } |
| 21060 | } |
| 21061 | |
| 21062 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_eq_8_subtile_m) { |
| 21063 | TEST_REQUIRES_ARM_NEON; |
| 21064 | for (uint32_t m = 1; m <= 2; m++) { |
| 21065 | GemmMicrokernelTester() |
| 21066 | .mr(2) |
| 21067 | .nr(16) |
| 21068 | .kr(1) |
| 21069 | .sr(1) |
| 21070 | .m(m) |
| 21071 | .n(16) |
| 21072 | .k(8) |
| 21073 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21074 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21075 | } |
| 21076 | } |
| 21077 | |
| 21078 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_eq_8_subtile_n) { |
| 21079 | TEST_REQUIRES_ARM_NEON; |
| 21080 | for (uint32_t n = 1; n <= 16; n++) { |
| 21081 | GemmMicrokernelTester() |
| 21082 | .mr(2) |
| 21083 | .nr(16) |
| 21084 | .kr(1) |
| 21085 | .sr(1) |
| 21086 | .m(2) |
| 21087 | .n(n) |
| 21088 | .k(8) |
| 21089 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21090 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21091 | } |
| 21092 | } |
| 21093 | |
| 21094 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_lt_8) { |
| 21095 | TEST_REQUIRES_ARM_NEON; |
| 21096 | for (size_t k = 1; k < 8; k++) { |
| 21097 | GemmMicrokernelTester() |
| 21098 | .mr(2) |
| 21099 | .nr(16) |
| 21100 | .kr(1) |
| 21101 | .sr(1) |
| 21102 | .m(2) |
| 21103 | .n(16) |
| 21104 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21105 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21106 | } |
| 21107 | } |
| 21108 | |
| 21109 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_lt_8_strided_a) { |
| 21110 | TEST_REQUIRES_ARM_NEON; |
| 21111 | for (size_t k = 1; k < 8; k++) { |
| 21112 | GemmMicrokernelTester() |
| 21113 | .mr(2) |
| 21114 | .nr(16) |
| 21115 | .kr(1) |
| 21116 | .sr(1) |
| 21117 | .m(2) |
| 21118 | .n(16) |
| 21119 | .k(k) |
| 21120 | .a_stride(11) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21121 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21122 | } |
| 21123 | } |
| 21124 | |
| 21125 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_lt_8_subtile) { |
| 21126 | TEST_REQUIRES_ARM_NEON; |
| 21127 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21128 | for (uint32_t n = 1; n <= 16; n++) { |
| 21129 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21130 | GemmMicrokernelTester() |
| 21131 | .mr(2) |
| 21132 | .nr(16) |
| 21133 | .kr(1) |
| 21134 | .sr(1) |
| 21135 | .m(m) |
| 21136 | .n(n) |
| 21137 | .k(k) |
| 21138 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21139 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21140 | } |
| 21141 | } |
| 21142 | } |
| 21143 | } |
| 21144 | |
| 21145 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_gt_8) { |
| 21146 | TEST_REQUIRES_ARM_NEON; |
| 21147 | for (size_t k = 9; k < 16; k++) { |
| 21148 | GemmMicrokernelTester() |
| 21149 | .mr(2) |
| 21150 | .nr(16) |
| 21151 | .kr(1) |
| 21152 | .sr(1) |
| 21153 | .m(2) |
| 21154 | .n(16) |
| 21155 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21156 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21157 | } |
| 21158 | } |
| 21159 | |
| 21160 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_gt_8_strided_a) { |
| 21161 | TEST_REQUIRES_ARM_NEON; |
| 21162 | for (size_t k = 9; k < 16; k++) { |
| 21163 | GemmMicrokernelTester() |
| 21164 | .mr(2) |
| 21165 | .nr(16) |
| 21166 | .kr(1) |
| 21167 | .sr(1) |
| 21168 | .m(2) |
| 21169 | .n(16) |
| 21170 | .k(k) |
| 21171 | .a_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21172 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21173 | } |
| 21174 | } |
| 21175 | |
| 21176 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_gt_8_subtile) { |
| 21177 | TEST_REQUIRES_ARM_NEON; |
| 21178 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21179 | for (uint32_t n = 1; n <= 16; n++) { |
| 21180 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21181 | GemmMicrokernelTester() |
| 21182 | .mr(2) |
| 21183 | .nr(16) |
| 21184 | .kr(1) |
| 21185 | .sr(1) |
| 21186 | .m(m) |
| 21187 | .n(n) |
| 21188 | .k(k) |
| 21189 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21190 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21191 | } |
| 21192 | } |
| 21193 | } |
| 21194 | } |
| 21195 | |
| 21196 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_div_8) { |
| 21197 | TEST_REQUIRES_ARM_NEON; |
| 21198 | for (size_t k = 16; k <= 80; k += 8) { |
| 21199 | GemmMicrokernelTester() |
| 21200 | .mr(2) |
| 21201 | .nr(16) |
| 21202 | .kr(1) |
| 21203 | .sr(1) |
| 21204 | .m(2) |
| 21205 | .n(16) |
| 21206 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21207 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21208 | } |
| 21209 | } |
| 21210 | |
| 21211 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_div_8_strided_a) { |
| 21212 | TEST_REQUIRES_ARM_NEON; |
| 21213 | for (size_t k = 16; k <= 80; k += 8) { |
| 21214 | GemmMicrokernelTester() |
| 21215 | .mr(2) |
| 21216 | .nr(16) |
| 21217 | .kr(1) |
| 21218 | .sr(1) |
| 21219 | .m(2) |
| 21220 | .n(16) |
| 21221 | .k(k) |
| 21222 | .a_stride(83) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21223 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21224 | } |
| 21225 | } |
| 21226 | |
| 21227 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, k_div_8_subtile) { |
| 21228 | TEST_REQUIRES_ARM_NEON; |
| 21229 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21230 | for (uint32_t n = 1; n <= 16; n++) { |
| 21231 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21232 | GemmMicrokernelTester() |
| 21233 | .mr(2) |
| 21234 | .nr(16) |
| 21235 | .kr(1) |
| 21236 | .sr(1) |
| 21237 | .m(m) |
| 21238 | .n(n) |
| 21239 | .k(k) |
| 21240 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21241 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21242 | } |
| 21243 | } |
| 21244 | } |
| 21245 | } |
| 21246 | |
| 21247 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_gt_16) { |
| 21248 | TEST_REQUIRES_ARM_NEON; |
| 21249 | for (uint32_t n = 17; n < 32; n++) { |
| 21250 | for (size_t k = 1; k <= 40; k += 9) { |
| 21251 | GemmMicrokernelTester() |
| 21252 | .mr(2) |
| 21253 | .nr(16) |
| 21254 | .kr(1) |
| 21255 | .sr(1) |
| 21256 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 21257 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21258 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21259 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21260 | } |
| 21261 | } |
| 21262 | } |
| 21263 | |
| 21264 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_gt_16_strided_cn) { |
| 21265 | TEST_REQUIRES_ARM_NEON; |
| 21266 | for (uint32_t n = 17; n < 32; n++) { |
| 21267 | for (size_t k = 1; k <= 40; k += 9) { |
| 21268 | GemmMicrokernelTester() |
| 21269 | .mr(2) |
| 21270 | .nr(16) |
| 21271 | .kr(1) |
| 21272 | .sr(1) |
| 21273 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 21274 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21275 | .k(k) |
| 21276 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21277 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21278 | } |
| 21279 | } |
| 21280 | } |
| 21281 | |
| 21282 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_gt_16_strided_a) { |
| 21283 | TEST_REQUIRES_ARM_NEON; |
| 21284 | for (uint32_t n = 17; n < 32; n++) { |
| 21285 | for (size_t k = 1; k <= 40; k += 9) { |
| 21286 | GemmMicrokernelTester() |
| 21287 | .mr(2) |
| 21288 | .nr(16) |
| 21289 | .kr(1) |
| 21290 | .sr(1) |
| 21291 | .m(2) |
| 21292 | .n(n) |
| 21293 | .k(k) |
| 21294 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21295 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21296 | } |
| 21297 | } |
| 21298 | } |
| 21299 | |
| 21300 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_gt_16_subtile) { |
| 21301 | TEST_REQUIRES_ARM_NEON; |
| 21302 | for (uint32_t n = 17; n < 32; n++) { |
| 21303 | for (size_t k = 1; k <= 40; k += 9) { |
| 21304 | for (uint32_t m = 1; m <= 2; m++) { |
| 21305 | GemmMicrokernelTester() |
| 21306 | .mr(2) |
| 21307 | .nr(16) |
| 21308 | .kr(1) |
| 21309 | .sr(1) |
| 21310 | .m(m) |
| 21311 | .n(n) |
| 21312 | .k(k) |
| 21313 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21314 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21315 | } |
| 21316 | } |
| 21317 | } |
| 21318 | } |
| 21319 | |
| 21320 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_div_16) { |
| 21321 | TEST_REQUIRES_ARM_NEON; |
| 21322 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 21323 | for (size_t k = 1; k <= 40; k += 9) { |
| 21324 | GemmMicrokernelTester() |
| 21325 | .mr(2) |
| 21326 | .nr(16) |
| 21327 | .kr(1) |
| 21328 | .sr(1) |
| 21329 | .m(2) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 21330 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21331 | .k(k) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21332 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21333 | } |
| 21334 | } |
| 21335 | } |
| 21336 | |
| 21337 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_div_16_strided_cn) { |
| 21338 | TEST_REQUIRES_ARM_NEON; |
| 21339 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 21340 | for (size_t k = 1; k <= 40; k += 9) { |
| 21341 | GemmMicrokernelTester() |
| 21342 | .mr(2) |
| 21343 | .nr(16) |
| 21344 | .kr(1) |
| 21345 | .sr(1) |
| 21346 | .m(2) |
| 21347 | .n(n) |
| 21348 | .k(k) |
| 21349 | .cn_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21350 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21351 | } |
| 21352 | } |
| 21353 | } |
| 21354 | |
| 21355 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_div_16_strided_a) { |
| 21356 | TEST_REQUIRES_ARM_NEON; |
| 21357 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 21358 | for (size_t k = 1; k <= 40; k += 9) { |
| 21359 | GemmMicrokernelTester() |
| 21360 | .mr(2) |
| 21361 | .nr(16) |
| 21362 | .kr(1) |
| 21363 | .sr(1) |
| 21364 | .m(2) |
| 21365 | .n(n) |
| 21366 | .k(k) |
| 21367 | .a_stride(43) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21368 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21369 | } |
| 21370 | } |
| 21371 | } |
| 21372 | |
| 21373 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, n_div_16_subtile) { |
| 21374 | TEST_REQUIRES_ARM_NEON; |
| 21375 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 21376 | for (size_t k = 1; k <= 40; k += 9) { |
| 21377 | for (uint32_t m = 1; m <= 2; m++) { |
| 21378 | GemmMicrokernelTester() |
| 21379 | .mr(2) |
| 21380 | .nr(16) |
| 21381 | .kr(1) |
| 21382 | .sr(1) |
| 21383 | .m(m) |
| 21384 | .n(n) |
| 21385 | .k(k) |
| 21386 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21387 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21388 | } |
| 21389 | } |
| 21390 | } |
| 21391 | } |
| 21392 | |
| 21393 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, strided_cm_subtile) { |
| 21394 | TEST_REQUIRES_ARM_NEON; |
| 21395 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21396 | for (uint32_t n = 1; n <= 16; n++) { |
| 21397 | for (uint32_t m = 1; m <= 2; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21398 | GemmMicrokernelTester() |
| 21399 | .mr(2) |
| 21400 | .nr(16) |
| 21401 | .kr(1) |
| 21402 | .sr(1) |
| 21403 | .m(m) |
| 21404 | .n(n) |
| 21405 | .k(k) |
| 21406 | .cm_stride(19) |
| 21407 | .iterations(1) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21408 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21409 | } |
| 21410 | } |
| 21411 | } |
| 21412 | } |
| 21413 | |
| 21414 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, qmin) { |
| 21415 | TEST_REQUIRES_ARM_NEON; |
| 21416 | GemmMicrokernelTester() |
| 21417 | .mr(2) |
| 21418 | .nr(16) |
| 21419 | .kr(1) |
| 21420 | .sr(1) |
| 21421 | .m(2) |
| 21422 | .n(16) |
| 21423 | .k(8) |
| 21424 | .qmin(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21425 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21426 | } |
| 21427 | |
| 21428 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, qmax) { |
| 21429 | TEST_REQUIRES_ARM_NEON; |
| 21430 | GemmMicrokernelTester() |
| 21431 | .mr(2) |
| 21432 | .nr(16) |
| 21433 | .kr(1) |
| 21434 | .sr(1) |
| 21435 | .m(2) |
| 21436 | .n(16) |
| 21437 | .k(8) |
| 21438 | .qmax(128) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21439 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21440 | } |
| 21441 | |
| 21442 | TEST(QS8_GEMM_MINMAX_RNDNU_2X16__NEON_MULL_ADDW_DUP, strided_cm) { |
| 21443 | TEST_REQUIRES_ARM_NEON; |
| 21444 | GemmMicrokernelTester() |
| 21445 | .mr(2) |
| 21446 | .nr(16) |
| 21447 | .kr(1) |
| 21448 | .sr(1) |
| 21449 | .m(2) |
| 21450 | .n(16) |
| 21451 | .k(8) |
| 21452 | .cm_stride(19) |
Marat Dukhan | 50323b8 | 2022-01-11 00:12:01 -0800 | [diff] [blame] | 21453 | .Test(xnn_qs8_gemm_minmax_rndnu_ukernel_2x16__neon_mull_addw_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21454 | } |
| 21455 | #endif // XNN_ARCH_ARM || XNN_ARCH_ARM64 |
| 21456 | |
| 21457 | |
| 21458 | #if XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY && XNN_PLATFORM_JIT |
| 21459 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8) { |
| 21460 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21461 | GemmMicrokernelTester() |
| 21462 | .mr(4) |
| 21463 | .nr(8) |
| 21464 | .kr(1) |
| 21465 | .sr(1) |
| 21466 | .m(4) |
| 21467 | .n(8) |
| 21468 | .k(8) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21469 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21470 | } |
| 21471 | |
| 21472 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, strided_cn) { |
| 21473 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21474 | GemmMicrokernelTester() |
| 21475 | .mr(4) |
| 21476 | .nr(8) |
| 21477 | .kr(1) |
| 21478 | .sr(1) |
| 21479 | .m(4) |
| 21480 | .n(8) |
| 21481 | .k(8) |
| 21482 | .cn_stride(11) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21483 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21484 | } |
| 21485 | |
| 21486 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_strided_a) { |
| 21487 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21488 | GemmMicrokernelTester() |
| 21489 | .mr(4) |
| 21490 | .nr(8) |
| 21491 | .kr(1) |
| 21492 | .sr(1) |
| 21493 | .m(4) |
| 21494 | .n(8) |
| 21495 | .k(8) |
| 21496 | .a_stride(11) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21497 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21498 | } |
| 21499 | |
| 21500 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile) { |
| 21501 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21502 | for (uint32_t n = 1; n <= 8; n++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21503 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21504 | GemmMicrokernelTester() |
| 21505 | .mr(4) |
| 21506 | .nr(8) |
| 21507 | .kr(1) |
| 21508 | .sr(1) |
| 21509 | .m(m) |
| 21510 | .n(n) |
| 21511 | .k(8) |
| 21512 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21513 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21514 | } |
| 21515 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21516 | } |
| 21517 | |
| 21518 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile_m) { |
| 21519 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21520 | for (uint32_t m = 1; m <= 4; m++) { |
| 21521 | GemmMicrokernelTester() |
| 21522 | .mr(4) |
| 21523 | .nr(8) |
| 21524 | .kr(1) |
| 21525 | .sr(1) |
| 21526 | .m(m) |
| 21527 | .n(8) |
| 21528 | .k(8) |
| 21529 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21530 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21531 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21532 | } |
| 21533 | |
| 21534 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_eq_8_subtile_n) { |
| 21535 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21536 | for (uint32_t n = 1; n <= 8; n++) { |
| 21537 | GemmMicrokernelTester() |
| 21538 | .mr(4) |
| 21539 | .nr(8) |
| 21540 | .kr(1) |
| 21541 | .sr(1) |
| 21542 | .m(4) |
| 21543 | .n(n) |
| 21544 | .k(8) |
| 21545 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21546 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21547 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21548 | } |
| 21549 | |
| 21550 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_lt_8) { |
| 21551 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21552 | for (size_t k = 1; k < 8; k++) { |
| 21553 | GemmMicrokernelTester() |
| 21554 | .mr(4) |
| 21555 | .nr(8) |
| 21556 | .kr(1) |
| 21557 | .sr(1) |
| 21558 | .m(4) |
| 21559 | .n(8) |
| 21560 | .k(k) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21561 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21562 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21563 | } |
| 21564 | |
| 21565 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_lt_8_strided_a) { |
| 21566 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21567 | for (size_t k = 1; k < 8; k++) { |
| 21568 | GemmMicrokernelTester() |
| 21569 | .mr(4) |
| 21570 | .nr(8) |
| 21571 | .kr(1) |
| 21572 | .sr(1) |
| 21573 | .m(4) |
| 21574 | .n(8) |
| 21575 | .k(k) |
| 21576 | .a_stride(11) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21577 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21578 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21579 | } |
| 21580 | |
| 21581 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_lt_8_subtile) { |
| 21582 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21583 | for (size_t k = 1; k < 8; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21584 | for (uint32_t n = 1; n <= 8; n++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21585 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21586 | GemmMicrokernelTester() |
| 21587 | .mr(4) |
| 21588 | .nr(8) |
| 21589 | .kr(1) |
| 21590 | .sr(1) |
| 21591 | .m(m) |
| 21592 | .n(n) |
| 21593 | .k(k) |
| 21594 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21595 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21596 | } |
| 21597 | } |
| 21598 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21599 | } |
| 21600 | |
| 21601 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_gt_8) { |
| 21602 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21603 | for (size_t k = 9; k < 16; k++) { |
| 21604 | GemmMicrokernelTester() |
| 21605 | .mr(4) |
| 21606 | .nr(8) |
| 21607 | .kr(1) |
| 21608 | .sr(1) |
| 21609 | .m(4) |
| 21610 | .n(8) |
| 21611 | .k(k) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21612 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21613 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21614 | } |
| 21615 | |
| 21616 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_gt_8_strided_a) { |
| 21617 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21618 | for (size_t k = 9; k < 16; k++) { |
| 21619 | GemmMicrokernelTester() |
| 21620 | .mr(4) |
| 21621 | .nr(8) |
| 21622 | .kr(1) |
| 21623 | .sr(1) |
| 21624 | .m(4) |
| 21625 | .n(8) |
| 21626 | .k(k) |
| 21627 | .a_stride(19) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21628 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21629 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21630 | } |
| 21631 | |
| 21632 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_gt_8_subtile) { |
| 21633 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21634 | for (size_t k = 9; k < 16; k++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21635 | for (uint32_t n = 1; n <= 8; n++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21636 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21637 | GemmMicrokernelTester() |
| 21638 | .mr(4) |
| 21639 | .nr(8) |
| 21640 | .kr(1) |
| 21641 | .sr(1) |
| 21642 | .m(m) |
| 21643 | .n(n) |
| 21644 | .k(k) |
| 21645 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21646 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21647 | } |
| 21648 | } |
| 21649 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21650 | } |
| 21651 | |
| 21652 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_div_8) { |
| 21653 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21654 | for (size_t k = 16; k <= 80; k += 8) { |
| 21655 | GemmMicrokernelTester() |
| 21656 | .mr(4) |
| 21657 | .nr(8) |
| 21658 | .kr(1) |
| 21659 | .sr(1) |
| 21660 | .m(4) |
| 21661 | .n(8) |
| 21662 | .k(k) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21663 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21664 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21665 | } |
| 21666 | |
| 21667 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_div_8_strided_a) { |
| 21668 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21669 | for (size_t k = 16; k <= 80; k += 8) { |
| 21670 | GemmMicrokernelTester() |
| 21671 | .mr(4) |
| 21672 | .nr(8) |
| 21673 | .kr(1) |
| 21674 | .sr(1) |
| 21675 | .m(4) |
| 21676 | .n(8) |
| 21677 | .k(k) |
| 21678 | .a_stride(83) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21679 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21680 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21681 | } |
| 21682 | |
| 21683 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, k_div_8_subtile) { |
| 21684 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21685 | for (size_t k = 16; k <= 80; k += 8) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21686 | for (uint32_t n = 1; n <= 8; n++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21687 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21688 | GemmMicrokernelTester() |
| 21689 | .mr(4) |
| 21690 | .nr(8) |
| 21691 | .kr(1) |
| 21692 | .sr(1) |
| 21693 | .m(m) |
| 21694 | .n(n) |
| 21695 | .k(k) |
| 21696 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21697 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21698 | } |
| 21699 | } |
| 21700 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21701 | } |
| 21702 | |
| 21703 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8) { |
| 21704 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21705 | for (uint32_t n = 9; n < 16; n++) { |
| 21706 | for (size_t k = 1; k <= 40; k += 9) { |
| 21707 | GemmMicrokernelTester() |
| 21708 | .mr(4) |
| 21709 | .nr(8) |
| 21710 | .kr(1) |
| 21711 | .sr(1) |
| 21712 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 21713 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21714 | .k(k) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21715 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21716 | } |
| 21717 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21718 | } |
| 21719 | |
| 21720 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8_strided_cn) { |
| 21721 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21722 | for (uint32_t n = 9; n < 16; n++) { |
| 21723 | for (size_t k = 1; k <= 40; k += 9) { |
| 21724 | GemmMicrokernelTester() |
| 21725 | .mr(4) |
| 21726 | .nr(8) |
| 21727 | .kr(1) |
| 21728 | .sr(1) |
| 21729 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 21730 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21731 | .k(k) |
| 21732 | .cn_stride(11) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21733 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21734 | } |
| 21735 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21736 | } |
| 21737 | |
| 21738 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8_strided_a) { |
| 21739 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21740 | for (uint32_t n = 9; n < 16; n++) { |
| 21741 | for (size_t k = 1; k <= 40; k += 9) { |
| 21742 | GemmMicrokernelTester() |
| 21743 | .mr(4) |
| 21744 | .nr(8) |
| 21745 | .kr(1) |
| 21746 | .sr(1) |
| 21747 | .m(4) |
| 21748 | .n(n) |
| 21749 | .k(k) |
| 21750 | .a_stride(43) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21751 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21752 | } |
| 21753 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21754 | } |
| 21755 | |
| 21756 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_gt_8_subtile) { |
| 21757 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21758 | for (uint32_t n = 9; n < 16; n++) { |
| 21759 | for (size_t k = 1; k <= 40; k += 9) { |
| 21760 | for (uint32_t m = 1; m <= 4; m++) { |
| 21761 | GemmMicrokernelTester() |
| 21762 | .mr(4) |
| 21763 | .nr(8) |
| 21764 | .kr(1) |
| 21765 | .sr(1) |
| 21766 | .m(m) |
| 21767 | .n(n) |
| 21768 | .k(k) |
| 21769 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21770 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21771 | } |
| 21772 | } |
| 21773 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21774 | } |
| 21775 | |
| 21776 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8) { |
| 21777 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21778 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 21779 | for (size_t k = 1; k <= 40; k += 9) { |
| 21780 | GemmMicrokernelTester() |
| 21781 | .mr(4) |
| 21782 | .nr(8) |
| 21783 | .kr(1) |
| 21784 | .sr(1) |
| 21785 | .m(4) |
Zhi An Ng | af9ff85 | 2022-01-13 10:48:37 -0800 | [diff] [blame] | 21786 | .n(n) |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21787 | .k(k) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21788 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21789 | } |
| 21790 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21791 | } |
| 21792 | |
| 21793 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8_strided_cn) { |
| 21794 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21795 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 21796 | for (size_t k = 1; k <= 40; k += 9) { |
| 21797 | GemmMicrokernelTester() |
| 21798 | .mr(4) |
| 21799 | .nr(8) |
| 21800 | .kr(1) |
| 21801 | .sr(1) |
| 21802 | .m(4) |
| 21803 | .n(n) |
| 21804 | .k(k) |
| 21805 | .cn_stride(11) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21806 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21807 | } |
| 21808 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21809 | } |
| 21810 | |
| 21811 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8_strided_a) { |
| 21812 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21813 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 21814 | for (size_t k = 1; k <= 40; k += 9) { |
| 21815 | GemmMicrokernelTester() |
| 21816 | .mr(4) |
| 21817 | .nr(8) |
| 21818 | .kr(1) |
| 21819 | .sr(1) |
| 21820 | .m(4) |
| 21821 | .n(n) |
| 21822 | .k(k) |
| 21823 | .a_stride(43) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21824 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21825 | } |
| 21826 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21827 | } |
| 21828 | |
| 21829 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, n_div_8_subtile) { |
| 21830 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21831 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 21832 | for (size_t k = 1; k <= 40; k += 9) { |
| 21833 | for (uint32_t m = 1; m <= 4; m++) { |
| 21834 | GemmMicrokernelTester() |
| 21835 | .mr(4) |
| 21836 | .nr(8) |
| 21837 | .kr(1) |
| 21838 | .sr(1) |
| 21839 | .m(m) |
| 21840 | .n(n) |
| 21841 | .k(k) |
| 21842 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21843 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21844 | } |
| 21845 | } |
| 21846 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21847 | } |
| 21848 | |
| 21849 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, strided_cm_subtile) { |
| 21850 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21851 | for (size_t k = 1; k <= 40; k += 9) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21852 | for (uint32_t n = 1; n <= 8; n++) { |
Zhi An Ng | 83844ae | 2022-01-14 09:52:25 -0800 | [diff] [blame] | 21853 | for (uint32_t m = 1; m <= 4; m++) { |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21854 | GemmMicrokernelTester() |
| 21855 | .mr(4) |
| 21856 | .nr(8) |
| 21857 | .kr(1) |
| 21858 | .sr(1) |
| 21859 | .m(m) |
| 21860 | .n(n) |
| 21861 | .k(k) |
| 21862 | .cm_stride(11) |
| 21863 | .iterations(1) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21864 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21865 | } |
| 21866 | } |
| 21867 | } |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21868 | } |
| 21869 | |
| 21870 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, qmin) { |
| 21871 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21872 | GemmMicrokernelTester() |
| 21873 | .mr(4) |
| 21874 | .nr(8) |
| 21875 | .kr(1) |
| 21876 | .sr(1) |
| 21877 | .m(4) |
| 21878 | .n(8) |
| 21879 | .k(8) |
| 21880 | .qmin(128) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21881 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21882 | } |
| 21883 | |
| 21884 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, qmax) { |
| 21885 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21886 | GemmMicrokernelTester() |
| 21887 | .mr(4) |
| 21888 | .nr(8) |
| 21889 | .kr(1) |
| 21890 | .sr(1) |
| 21891 | .m(4) |
| 21892 | .n(8) |
| 21893 | .k(8) |
| 21894 | .qmax(128) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21895 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21896 | } |
| 21897 | |
| 21898 | TEST(GENERATE_QS8_GEMM_RNDNU_4X8__AARCH32_NEON_MLAL_LANE_PRFM_LD64, strided_cm) { |
| 21899 | TEST_REQUIRES_ARM_NEON; |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21900 | GemmMicrokernelTester() |
| 21901 | .mr(4) |
| 21902 | .nr(8) |
| 21903 | .kr(1) |
| 21904 | .sr(1) |
| 21905 | .m(4) |
| 21906 | .n(8) |
| 21907 | .k(8) |
| 21908 | .cm_stride(11) |
Zhi An Ng | 0ec25cf | 2022-01-19 11:38:55 -0800 | [diff] [blame] | 21909 | .Test(xnn_generate_qs8_gemm_rndnu_ukernel_4x8__aarch32_neon_mlal_lane_prfm_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu); |
Zhi An Ng | 4c1fd6f | 2022-01-10 19:35:06 -0800 | [diff] [blame] | 21910 | } |
| 21911 | #endif // XNN_ARCH_ARM && XNN_ENABLE_ASSEMBLY && XNN_PLATFORM_JIT |