XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | // |
| 9 | // Auto-generated file. Do not edit! |
Marat Dukhan | 163a7e6 | 2020-04-09 04:19:26 -0700 | [diff] [blame] | 10 | // Specification: test/f16-gemm-minmax.yaml |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 11 | // Generator: tools/generate-gemm-test.py |
| 12 | |
| 13 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 14 | #include <gtest/gtest.h> |
| 15 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 16 | #include <xnnpack/common.h> |
| 17 | #include <xnnpack/isa-checks.h> |
| 18 | |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 19 | #include <xnnpack/gemm.h> |
| 20 | #include <xnnpack/igemm.h> |
| 21 | #include <xnnpack/ppmm.h> |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 22 | #include "gemm-microkernel-tester.h" |
| 23 | |
| 24 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 25 | #if XNN_ARCH_ARM64 |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 26 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 27 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 28 | GemmMicrokernelTester() |
| 29 | .mr(4) |
| 30 | .nr(8) |
| 31 | .kr(1) |
| 32 | .sr(1) |
| 33 | .m(4) |
| 34 | .n(8) |
| 35 | .k(4) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 36 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 37 | } |
| 38 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 39 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 40 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 41 | GemmMicrokernelTester() |
| 42 | .mr(4) |
| 43 | .nr(8) |
| 44 | .kr(1) |
| 45 | .sr(1) |
| 46 | .m(4) |
| 47 | .n(8) |
| 48 | .k(4) |
| 49 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 50 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 51 | } |
| 52 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 53 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 54 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 55 | GemmMicrokernelTester() |
| 56 | .mr(4) |
| 57 | .nr(8) |
| 58 | .kr(1) |
| 59 | .sr(1) |
| 60 | .m(4) |
| 61 | .n(8) |
| 62 | .k(4) |
| 63 | .a_stride(7) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 64 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 65 | } |
| 66 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 67 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 68 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 69 | for (uint32_t m = 1; m <= 4; m++) { |
| 70 | for (uint32_t n = 1; n <= 8; n++) { |
| 71 | GemmMicrokernelTester() |
| 72 | .mr(4) |
| 73 | .nr(8) |
| 74 | .kr(1) |
| 75 | .sr(1) |
| 76 | .m(m) |
| 77 | .n(n) |
| 78 | .k(4) |
| 79 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 80 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 81 | } |
| 82 | } |
| 83 | } |
| 84 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 85 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 86 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 87 | for (uint32_t m = 1; m <= 4; m++) { |
| 88 | GemmMicrokernelTester() |
| 89 | .mr(4) |
| 90 | .nr(8) |
| 91 | .kr(1) |
| 92 | .sr(1) |
| 93 | .m(m) |
| 94 | .n(8) |
| 95 | .k(4) |
| 96 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 97 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 98 | } |
| 99 | } |
| 100 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 101 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 102 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 103 | for (uint32_t n = 1; n <= 8; n++) { |
| 104 | GemmMicrokernelTester() |
| 105 | .mr(4) |
| 106 | .nr(8) |
| 107 | .kr(1) |
| 108 | .sr(1) |
| 109 | .m(4) |
| 110 | .n(n) |
| 111 | .k(4) |
| 112 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 113 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 114 | } |
| 115 | } |
| 116 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 117 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_lt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 118 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 119 | for (size_t k = 1; k < 4; k++) { |
| 120 | GemmMicrokernelTester() |
| 121 | .mr(4) |
| 122 | .nr(8) |
| 123 | .kr(1) |
| 124 | .sr(1) |
| 125 | .m(4) |
| 126 | .n(8) |
| 127 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 128 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 129 | } |
| 130 | } |
| 131 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 132 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_lt_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 133 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 134 | for (size_t k = 1; k < 4; k++) { |
| 135 | GemmMicrokernelTester() |
| 136 | .mr(4) |
| 137 | .nr(8) |
| 138 | .kr(1) |
| 139 | .sr(1) |
| 140 | .m(4) |
| 141 | .n(8) |
| 142 | .k(k) |
| 143 | .a_stride(7) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 144 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 145 | } |
| 146 | } |
| 147 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 148 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 149 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 150 | for (size_t k = 1; k < 4; k++) { |
| 151 | for (uint32_t m = 1; m <= 4; m++) { |
| 152 | for (uint32_t n = 1; n <= 8; n++) { |
| 153 | GemmMicrokernelTester() |
| 154 | .mr(4) |
| 155 | .nr(8) |
| 156 | .kr(1) |
| 157 | .sr(1) |
| 158 | .m(m) |
| 159 | .n(n) |
| 160 | .k(k) |
| 161 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 162 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 163 | } |
| 164 | } |
| 165 | } |
| 166 | } |
| 167 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 168 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_gt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 169 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 170 | for (size_t k = 5; k < 8; k++) { |
| 171 | GemmMicrokernelTester() |
| 172 | .mr(4) |
| 173 | .nr(8) |
| 174 | .kr(1) |
| 175 | .sr(1) |
| 176 | .m(4) |
| 177 | .n(8) |
| 178 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 179 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 180 | } |
| 181 | } |
| 182 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 183 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_gt_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 184 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 185 | for (size_t k = 5; k < 8; k++) { |
| 186 | GemmMicrokernelTester() |
| 187 | .mr(4) |
| 188 | .nr(8) |
| 189 | .kr(1) |
| 190 | .sr(1) |
| 191 | .m(4) |
| 192 | .n(8) |
| 193 | .k(k) |
| 194 | .a_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 195 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 196 | } |
| 197 | } |
| 198 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 199 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 200 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 201 | for (size_t k = 5; k < 8; k++) { |
| 202 | for (uint32_t m = 1; m <= 4; m++) { |
| 203 | for (uint32_t n = 1; n <= 8; n++) { |
| 204 | GemmMicrokernelTester() |
| 205 | .mr(4) |
| 206 | .nr(8) |
| 207 | .kr(1) |
| 208 | .sr(1) |
| 209 | .m(m) |
| 210 | .n(n) |
| 211 | .k(k) |
| 212 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 213 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 214 | } |
| 215 | } |
| 216 | } |
| 217 | } |
| 218 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 219 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_div_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 220 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 221 | for (size_t k = 8; k <= 40; k += 4) { |
| 222 | GemmMicrokernelTester() |
| 223 | .mr(4) |
| 224 | .nr(8) |
| 225 | .kr(1) |
| 226 | .sr(1) |
| 227 | .m(4) |
| 228 | .n(8) |
| 229 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 230 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 231 | } |
| 232 | } |
| 233 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 234 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_div_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 235 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 236 | for (size_t k = 8; k <= 40; k += 4) { |
| 237 | GemmMicrokernelTester() |
| 238 | .mr(4) |
| 239 | .nr(8) |
| 240 | .kr(1) |
| 241 | .sr(1) |
| 242 | .m(4) |
| 243 | .n(8) |
| 244 | .k(k) |
| 245 | .a_stride(43) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 246 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 247 | } |
| 248 | } |
| 249 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 250 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 251 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 252 | for (size_t k = 8; k <= 40; k += 4) { |
| 253 | for (uint32_t m = 1; m <= 4; m++) { |
| 254 | for (uint32_t n = 1; n <= 8; n++) { |
| 255 | GemmMicrokernelTester() |
| 256 | .mr(4) |
| 257 | .nr(8) |
| 258 | .kr(1) |
| 259 | .sr(1) |
| 260 | .m(m) |
| 261 | .n(n) |
| 262 | .k(k) |
| 263 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 264 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 265 | } |
| 266 | } |
| 267 | } |
| 268 | } |
| 269 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 270 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 271 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 272 | for (uint32_t n = 9; n < 16; n++) { |
| 273 | for (size_t k = 1; k <= 20; k += 5) { |
| 274 | GemmMicrokernelTester() |
| 275 | .mr(4) |
| 276 | .nr(8) |
| 277 | .kr(1) |
| 278 | .sr(1) |
| 279 | .m(4) |
| 280 | .n(8) |
| 281 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 282 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 283 | } |
| 284 | } |
| 285 | } |
| 286 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 287 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 288 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 289 | for (uint32_t n = 9; n < 16; n++) { |
| 290 | for (size_t k = 1; k <= 20; k += 5) { |
| 291 | GemmMicrokernelTester() |
| 292 | .mr(4) |
| 293 | .nr(8) |
| 294 | .kr(1) |
| 295 | .sr(1) |
| 296 | .m(4) |
| 297 | .n(8) |
| 298 | .k(k) |
| 299 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 300 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 301 | } |
| 302 | } |
| 303 | } |
| 304 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 305 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 306 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 307 | for (uint32_t n = 9; n < 16; n++) { |
| 308 | for (size_t k = 1; k <= 20; k += 5) { |
| 309 | GemmMicrokernelTester() |
| 310 | .mr(4) |
| 311 | .nr(8) |
| 312 | .kr(1) |
| 313 | .sr(1) |
| 314 | .m(4) |
| 315 | .n(n) |
| 316 | .k(k) |
| 317 | .a_stride(23) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 318 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 319 | } |
| 320 | } |
| 321 | } |
| 322 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 323 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 324 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 325 | for (uint32_t n = 9; n < 16; n++) { |
| 326 | for (size_t k = 1; k <= 20; k += 5) { |
| 327 | for (uint32_t m = 1; m <= 4; m++) { |
| 328 | GemmMicrokernelTester() |
| 329 | .mr(4) |
| 330 | .nr(8) |
| 331 | .kr(1) |
| 332 | .sr(1) |
| 333 | .m(m) |
| 334 | .n(n) |
| 335 | .k(k) |
| 336 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 337 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 338 | } |
| 339 | } |
| 340 | } |
| 341 | } |
| 342 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 343 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 344 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 345 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 346 | for (size_t k = 1; k <= 20; k += 5) { |
| 347 | GemmMicrokernelTester() |
| 348 | .mr(4) |
| 349 | .nr(8) |
| 350 | .kr(1) |
| 351 | .sr(1) |
| 352 | .m(4) |
| 353 | .n(8) |
| 354 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 355 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 356 | } |
| 357 | } |
| 358 | } |
| 359 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 360 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 361 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 362 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 363 | for (size_t k = 1; k <= 20; k += 5) { |
| 364 | GemmMicrokernelTester() |
| 365 | .mr(4) |
| 366 | .nr(8) |
| 367 | .kr(1) |
| 368 | .sr(1) |
| 369 | .m(4) |
| 370 | .n(n) |
| 371 | .k(k) |
| 372 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 373 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 374 | } |
| 375 | } |
| 376 | } |
| 377 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 378 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 379 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 380 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 381 | for (size_t k = 1; k <= 20; k += 5) { |
| 382 | GemmMicrokernelTester() |
| 383 | .mr(4) |
| 384 | .nr(8) |
| 385 | .kr(1) |
| 386 | .sr(1) |
| 387 | .m(4) |
| 388 | .n(n) |
| 389 | .k(k) |
| 390 | .a_stride(23) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 391 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 392 | } |
| 393 | } |
| 394 | } |
| 395 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 396 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 397 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 398 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 399 | for (size_t k = 1; k <= 20; k += 5) { |
| 400 | for (uint32_t m = 1; m <= 4; m++) { |
| 401 | GemmMicrokernelTester() |
| 402 | .mr(4) |
| 403 | .nr(8) |
| 404 | .kr(1) |
| 405 | .sr(1) |
| 406 | .m(m) |
| 407 | .n(n) |
| 408 | .k(k) |
| 409 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 410 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 411 | } |
| 412 | } |
| 413 | } |
| 414 | } |
| 415 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 416 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 417 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 418 | for (size_t k = 1; k <= 20; k += 5) { |
| 419 | for (uint32_t m = 1; m <= 4; m++) { |
| 420 | for (uint32_t n = 1; n <= 8; n++) { |
| 421 | GemmMicrokernelTester() |
| 422 | .mr(4) |
| 423 | .nr(8) |
| 424 | .kr(1) |
| 425 | .sr(1) |
| 426 | .m(m) |
| 427 | .n(n) |
| 428 | .k(k) |
| 429 | .cm_stride(11) |
| 430 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 431 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 432 | } |
| 433 | } |
| 434 | } |
| 435 | } |
| 436 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 437 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, qmin) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 438 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 439 | GemmMicrokernelTester() |
| 440 | .mr(4) |
| 441 | .nr(8) |
| 442 | .kr(1) |
| 443 | .sr(1) |
| 444 | .m(4) |
| 445 | .n(8) |
| 446 | .k(4) |
| 447 | .qmin(128) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 448 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 449 | } |
| 450 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 451 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, qmax) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 452 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 453 | GemmMicrokernelTester() |
| 454 | .mr(4) |
| 455 | .nr(8) |
| 456 | .kr(1) |
| 457 | .sr(1) |
| 458 | .m(4) |
| 459 | .n(8) |
| 460 | .k(4) |
| 461 | .qmax(128) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 462 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 463 | } |
| 464 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 465 | TEST(F16_GEMM_MINMAX_4X8__NEONFP16ARITH_LD64, strided_cm) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 466 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 467 | GemmMicrokernelTester() |
| 468 | .mr(4) |
| 469 | .nr(8) |
| 470 | .kr(1) |
| 471 | .sr(1) |
| 472 | .m(4) |
| 473 | .n(8) |
| 474 | .k(4) |
| 475 | .cm_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 476 | .Test(xnn_f16_gemm_minmax_ukernel_4x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 477 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 478 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 479 | |
| 480 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 481 | #if XNN_ARCH_ARM64 |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 482 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 483 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 484 | GemmMicrokernelTester() |
| 485 | .mr(6) |
| 486 | .nr(8) |
| 487 | .kr(1) |
| 488 | .sr(1) |
| 489 | .m(6) |
| 490 | .n(8) |
| 491 | .k(4) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 492 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 493 | } |
| 494 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 495 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 496 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 497 | GemmMicrokernelTester() |
| 498 | .mr(6) |
| 499 | .nr(8) |
| 500 | .kr(1) |
| 501 | .sr(1) |
| 502 | .m(6) |
| 503 | .n(8) |
| 504 | .k(4) |
| 505 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 506 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 507 | } |
| 508 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 509 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 510 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 511 | GemmMicrokernelTester() |
| 512 | .mr(6) |
| 513 | .nr(8) |
| 514 | .kr(1) |
| 515 | .sr(1) |
| 516 | .m(6) |
| 517 | .n(8) |
| 518 | .k(4) |
| 519 | .a_stride(7) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 520 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 521 | } |
| 522 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 523 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 524 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 525 | for (uint32_t m = 1; m <= 6; m++) { |
| 526 | for (uint32_t n = 1; n <= 8; n++) { |
| 527 | GemmMicrokernelTester() |
| 528 | .mr(6) |
| 529 | .nr(8) |
| 530 | .kr(1) |
| 531 | .sr(1) |
| 532 | .m(m) |
| 533 | .n(n) |
| 534 | .k(4) |
| 535 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 536 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 537 | } |
| 538 | } |
| 539 | } |
| 540 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 541 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 542 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 543 | for (uint32_t m = 1; m <= 6; m++) { |
| 544 | GemmMicrokernelTester() |
| 545 | .mr(6) |
| 546 | .nr(8) |
| 547 | .kr(1) |
| 548 | .sr(1) |
| 549 | .m(m) |
| 550 | .n(8) |
| 551 | .k(4) |
| 552 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 553 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 554 | } |
| 555 | } |
| 556 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 557 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 558 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 559 | for (uint32_t n = 1; n <= 8; n++) { |
| 560 | GemmMicrokernelTester() |
| 561 | .mr(6) |
| 562 | .nr(8) |
| 563 | .kr(1) |
| 564 | .sr(1) |
| 565 | .m(6) |
| 566 | .n(n) |
| 567 | .k(4) |
| 568 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 569 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 570 | } |
| 571 | } |
| 572 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 573 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_lt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 574 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 575 | for (size_t k = 1; k < 4; k++) { |
| 576 | GemmMicrokernelTester() |
| 577 | .mr(6) |
| 578 | .nr(8) |
| 579 | .kr(1) |
| 580 | .sr(1) |
| 581 | .m(6) |
| 582 | .n(8) |
| 583 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 584 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 585 | } |
| 586 | } |
| 587 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 588 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_lt_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 589 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 590 | for (size_t k = 1; k < 4; k++) { |
| 591 | GemmMicrokernelTester() |
| 592 | .mr(6) |
| 593 | .nr(8) |
| 594 | .kr(1) |
| 595 | .sr(1) |
| 596 | .m(6) |
| 597 | .n(8) |
| 598 | .k(k) |
| 599 | .a_stride(7) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 600 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 601 | } |
| 602 | } |
| 603 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 604 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 605 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 606 | for (size_t k = 1; k < 4; k++) { |
| 607 | for (uint32_t m = 1; m <= 6; m++) { |
| 608 | for (uint32_t n = 1; n <= 8; n++) { |
| 609 | GemmMicrokernelTester() |
| 610 | .mr(6) |
| 611 | .nr(8) |
| 612 | .kr(1) |
| 613 | .sr(1) |
| 614 | .m(m) |
| 615 | .n(n) |
| 616 | .k(k) |
| 617 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 618 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 619 | } |
| 620 | } |
| 621 | } |
| 622 | } |
| 623 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 624 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_gt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 625 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 626 | for (size_t k = 5; k < 8; k++) { |
| 627 | GemmMicrokernelTester() |
| 628 | .mr(6) |
| 629 | .nr(8) |
| 630 | .kr(1) |
| 631 | .sr(1) |
| 632 | .m(6) |
| 633 | .n(8) |
| 634 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 635 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 636 | } |
| 637 | } |
| 638 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 639 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_gt_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 640 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 641 | for (size_t k = 5; k < 8; k++) { |
| 642 | GemmMicrokernelTester() |
| 643 | .mr(6) |
| 644 | .nr(8) |
| 645 | .kr(1) |
| 646 | .sr(1) |
| 647 | .m(6) |
| 648 | .n(8) |
| 649 | .k(k) |
| 650 | .a_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 651 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 652 | } |
| 653 | } |
| 654 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 655 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 656 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 657 | for (size_t k = 5; k < 8; k++) { |
| 658 | for (uint32_t m = 1; m <= 6; m++) { |
| 659 | for (uint32_t n = 1; n <= 8; n++) { |
| 660 | GemmMicrokernelTester() |
| 661 | .mr(6) |
| 662 | .nr(8) |
| 663 | .kr(1) |
| 664 | .sr(1) |
| 665 | .m(m) |
| 666 | .n(n) |
| 667 | .k(k) |
| 668 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 669 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 670 | } |
| 671 | } |
| 672 | } |
| 673 | } |
| 674 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 675 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_div_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 676 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 677 | for (size_t k = 8; k <= 40; k += 4) { |
| 678 | GemmMicrokernelTester() |
| 679 | .mr(6) |
| 680 | .nr(8) |
| 681 | .kr(1) |
| 682 | .sr(1) |
| 683 | .m(6) |
| 684 | .n(8) |
| 685 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 686 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 687 | } |
| 688 | } |
| 689 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 690 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_div_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 691 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 692 | for (size_t k = 8; k <= 40; k += 4) { |
| 693 | GemmMicrokernelTester() |
| 694 | .mr(6) |
| 695 | .nr(8) |
| 696 | .kr(1) |
| 697 | .sr(1) |
| 698 | .m(6) |
| 699 | .n(8) |
| 700 | .k(k) |
| 701 | .a_stride(43) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 702 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 703 | } |
| 704 | } |
| 705 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 706 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 707 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 708 | for (size_t k = 8; k <= 40; k += 4) { |
| 709 | for (uint32_t m = 1; m <= 6; m++) { |
| 710 | for (uint32_t n = 1; n <= 8; n++) { |
| 711 | GemmMicrokernelTester() |
| 712 | .mr(6) |
| 713 | .nr(8) |
| 714 | .kr(1) |
| 715 | .sr(1) |
| 716 | .m(m) |
| 717 | .n(n) |
| 718 | .k(k) |
| 719 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 720 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 721 | } |
| 722 | } |
| 723 | } |
| 724 | } |
| 725 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 726 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 727 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 728 | for (uint32_t n = 9; n < 16; n++) { |
| 729 | for (size_t k = 1; k <= 20; k += 5) { |
| 730 | GemmMicrokernelTester() |
| 731 | .mr(6) |
| 732 | .nr(8) |
| 733 | .kr(1) |
| 734 | .sr(1) |
| 735 | .m(6) |
| 736 | .n(8) |
| 737 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 738 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 739 | } |
| 740 | } |
| 741 | } |
| 742 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 743 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 744 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 745 | for (uint32_t n = 9; n < 16; n++) { |
| 746 | for (size_t k = 1; k <= 20; k += 5) { |
| 747 | GemmMicrokernelTester() |
| 748 | .mr(6) |
| 749 | .nr(8) |
| 750 | .kr(1) |
| 751 | .sr(1) |
| 752 | .m(6) |
| 753 | .n(8) |
| 754 | .k(k) |
| 755 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 756 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 757 | } |
| 758 | } |
| 759 | } |
| 760 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 761 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 762 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 763 | for (uint32_t n = 9; n < 16; n++) { |
| 764 | for (size_t k = 1; k <= 20; k += 5) { |
| 765 | GemmMicrokernelTester() |
| 766 | .mr(6) |
| 767 | .nr(8) |
| 768 | .kr(1) |
| 769 | .sr(1) |
| 770 | .m(6) |
| 771 | .n(n) |
| 772 | .k(k) |
| 773 | .a_stride(23) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 774 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 775 | } |
| 776 | } |
| 777 | } |
| 778 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 779 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 780 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 781 | for (uint32_t n = 9; n < 16; n++) { |
| 782 | for (size_t k = 1; k <= 20; k += 5) { |
| 783 | for (uint32_t m = 1; m <= 6; m++) { |
| 784 | GemmMicrokernelTester() |
| 785 | .mr(6) |
| 786 | .nr(8) |
| 787 | .kr(1) |
| 788 | .sr(1) |
| 789 | .m(m) |
| 790 | .n(n) |
| 791 | .k(k) |
| 792 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 793 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 794 | } |
| 795 | } |
| 796 | } |
| 797 | } |
| 798 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 799 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 800 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 801 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 802 | for (size_t k = 1; k <= 20; k += 5) { |
| 803 | GemmMicrokernelTester() |
| 804 | .mr(6) |
| 805 | .nr(8) |
| 806 | .kr(1) |
| 807 | .sr(1) |
| 808 | .m(6) |
| 809 | .n(8) |
| 810 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 811 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 812 | } |
| 813 | } |
| 814 | } |
| 815 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 816 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 817 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 818 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 819 | for (size_t k = 1; k <= 20; k += 5) { |
| 820 | GemmMicrokernelTester() |
| 821 | .mr(6) |
| 822 | .nr(8) |
| 823 | .kr(1) |
| 824 | .sr(1) |
| 825 | .m(6) |
| 826 | .n(n) |
| 827 | .k(k) |
| 828 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 829 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 830 | } |
| 831 | } |
| 832 | } |
| 833 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 834 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 835 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 836 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 837 | for (size_t k = 1; k <= 20; k += 5) { |
| 838 | GemmMicrokernelTester() |
| 839 | .mr(6) |
| 840 | .nr(8) |
| 841 | .kr(1) |
| 842 | .sr(1) |
| 843 | .m(6) |
| 844 | .n(n) |
| 845 | .k(k) |
| 846 | .a_stride(23) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 847 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 848 | } |
| 849 | } |
| 850 | } |
| 851 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 852 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 853 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 854 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 855 | for (size_t k = 1; k <= 20; k += 5) { |
| 856 | for (uint32_t m = 1; m <= 6; m++) { |
| 857 | GemmMicrokernelTester() |
| 858 | .mr(6) |
| 859 | .nr(8) |
| 860 | .kr(1) |
| 861 | .sr(1) |
| 862 | .m(m) |
| 863 | .n(n) |
| 864 | .k(k) |
| 865 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 866 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 867 | } |
| 868 | } |
| 869 | } |
| 870 | } |
| 871 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 872 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 873 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 874 | for (size_t k = 1; k <= 20; k += 5) { |
| 875 | for (uint32_t m = 1; m <= 6; m++) { |
| 876 | for (uint32_t n = 1; n <= 8; n++) { |
| 877 | GemmMicrokernelTester() |
| 878 | .mr(6) |
| 879 | .nr(8) |
| 880 | .kr(1) |
| 881 | .sr(1) |
| 882 | .m(m) |
| 883 | .n(n) |
| 884 | .k(k) |
| 885 | .cm_stride(11) |
| 886 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 887 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 888 | } |
| 889 | } |
| 890 | } |
| 891 | } |
| 892 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 893 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, qmin) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 894 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 895 | GemmMicrokernelTester() |
| 896 | .mr(6) |
| 897 | .nr(8) |
| 898 | .kr(1) |
| 899 | .sr(1) |
| 900 | .m(6) |
| 901 | .n(8) |
| 902 | .k(4) |
| 903 | .qmin(128) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 904 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 905 | } |
| 906 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 907 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, qmax) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 908 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 909 | GemmMicrokernelTester() |
| 910 | .mr(6) |
| 911 | .nr(8) |
| 912 | .kr(1) |
| 913 | .sr(1) |
| 914 | .m(6) |
| 915 | .n(8) |
| 916 | .k(4) |
| 917 | .qmax(128) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 918 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 919 | } |
| 920 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 921 | TEST(F16_GEMM_MINMAX_6X8__NEONFP16ARITH_LD64, strided_cm) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 922 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 923 | GemmMicrokernelTester() |
| 924 | .mr(6) |
| 925 | .nr(8) |
| 926 | .kr(1) |
| 927 | .sr(1) |
| 928 | .m(6) |
| 929 | .n(8) |
| 930 | .k(4) |
| 931 | .cm_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 932 | .Test(xnn_f16_gemm_minmax_ukernel_6x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 933 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 934 | #endif // XNN_ARCH_ARM64 |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 935 | |
| 936 | |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 937 | #if XNN_ARCH_ARM64 |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 938 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 939 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 940 | GemmMicrokernelTester() |
| 941 | .mr(8) |
| 942 | .nr(8) |
| 943 | .kr(1) |
| 944 | .sr(1) |
| 945 | .m(8) |
| 946 | .n(8) |
| 947 | .k(4) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 948 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 949 | } |
| 950 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 951 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 952 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 953 | GemmMicrokernelTester() |
| 954 | .mr(8) |
| 955 | .nr(8) |
| 956 | .kr(1) |
| 957 | .sr(1) |
| 958 | .m(8) |
| 959 | .n(8) |
| 960 | .k(4) |
| 961 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 962 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 963 | } |
| 964 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 965 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 966 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 967 | GemmMicrokernelTester() |
| 968 | .mr(8) |
| 969 | .nr(8) |
| 970 | .kr(1) |
| 971 | .sr(1) |
| 972 | .m(8) |
| 973 | .n(8) |
| 974 | .k(4) |
| 975 | .a_stride(7) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 976 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 977 | } |
| 978 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 979 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 980 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 981 | for (uint32_t m = 1; m <= 8; m++) { |
| 982 | for (uint32_t n = 1; n <= 8; n++) { |
| 983 | GemmMicrokernelTester() |
| 984 | .mr(8) |
| 985 | .nr(8) |
| 986 | .kr(1) |
| 987 | .sr(1) |
| 988 | .m(m) |
| 989 | .n(n) |
| 990 | .k(4) |
| 991 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 992 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 993 | } |
| 994 | } |
| 995 | } |
| 996 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 997 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile_m) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 998 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 999 | for (uint32_t m = 1; m <= 8; m++) { |
| 1000 | GemmMicrokernelTester() |
| 1001 | .mr(8) |
| 1002 | .nr(8) |
| 1003 | .kr(1) |
| 1004 | .sr(1) |
| 1005 | .m(m) |
| 1006 | .n(8) |
| 1007 | .k(4) |
| 1008 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1009 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1010 | } |
| 1011 | } |
| 1012 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1013 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_eq_4_subtile_n) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1014 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1015 | for (uint32_t n = 1; n <= 8; n++) { |
| 1016 | GemmMicrokernelTester() |
| 1017 | .mr(8) |
| 1018 | .nr(8) |
| 1019 | .kr(1) |
| 1020 | .sr(1) |
| 1021 | .m(8) |
| 1022 | .n(n) |
| 1023 | .k(4) |
| 1024 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1025 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1026 | } |
| 1027 | } |
| 1028 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1029 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_lt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1030 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1031 | for (size_t k = 1; k < 4; k++) { |
| 1032 | GemmMicrokernelTester() |
| 1033 | .mr(8) |
| 1034 | .nr(8) |
| 1035 | .kr(1) |
| 1036 | .sr(1) |
| 1037 | .m(8) |
| 1038 | .n(8) |
| 1039 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1040 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1041 | } |
| 1042 | } |
| 1043 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1044 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_lt_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1045 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1046 | for (size_t k = 1; k < 4; k++) { |
| 1047 | GemmMicrokernelTester() |
| 1048 | .mr(8) |
| 1049 | .nr(8) |
| 1050 | .kr(1) |
| 1051 | .sr(1) |
| 1052 | .m(8) |
| 1053 | .n(8) |
| 1054 | .k(k) |
| 1055 | .a_stride(7) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1056 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1057 | } |
| 1058 | } |
| 1059 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1060 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_lt_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1061 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1062 | for (size_t k = 1; k < 4; k++) { |
| 1063 | for (uint32_t m = 1; m <= 8; m++) { |
| 1064 | for (uint32_t n = 1; n <= 8; n++) { |
| 1065 | GemmMicrokernelTester() |
| 1066 | .mr(8) |
| 1067 | .nr(8) |
| 1068 | .kr(1) |
| 1069 | .sr(1) |
| 1070 | .m(m) |
| 1071 | .n(n) |
| 1072 | .k(k) |
| 1073 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1074 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1075 | } |
| 1076 | } |
| 1077 | } |
| 1078 | } |
| 1079 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1080 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_gt_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1081 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1082 | for (size_t k = 5; k < 8; k++) { |
| 1083 | GemmMicrokernelTester() |
| 1084 | .mr(8) |
| 1085 | .nr(8) |
| 1086 | .kr(1) |
| 1087 | .sr(1) |
| 1088 | .m(8) |
| 1089 | .n(8) |
| 1090 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1091 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1092 | } |
| 1093 | } |
| 1094 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1095 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_gt_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1096 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1097 | for (size_t k = 5; k < 8; k++) { |
| 1098 | GemmMicrokernelTester() |
| 1099 | .mr(8) |
| 1100 | .nr(8) |
| 1101 | .kr(1) |
| 1102 | .sr(1) |
| 1103 | .m(8) |
| 1104 | .n(8) |
| 1105 | .k(k) |
| 1106 | .a_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1107 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1108 | } |
| 1109 | } |
| 1110 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1111 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_gt_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1112 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1113 | for (size_t k = 5; k < 8; k++) { |
| 1114 | for (uint32_t m = 1; m <= 8; m++) { |
| 1115 | for (uint32_t n = 1; n <= 8; n++) { |
| 1116 | GemmMicrokernelTester() |
| 1117 | .mr(8) |
| 1118 | .nr(8) |
| 1119 | .kr(1) |
| 1120 | .sr(1) |
| 1121 | .m(m) |
| 1122 | .n(n) |
| 1123 | .k(k) |
| 1124 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1125 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1126 | } |
| 1127 | } |
| 1128 | } |
| 1129 | } |
| 1130 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1131 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_div_4) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1132 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1133 | for (size_t k = 8; k <= 40; k += 4) { |
| 1134 | GemmMicrokernelTester() |
| 1135 | .mr(8) |
| 1136 | .nr(8) |
| 1137 | .kr(1) |
| 1138 | .sr(1) |
| 1139 | .m(8) |
| 1140 | .n(8) |
| 1141 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1142 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1143 | } |
| 1144 | } |
| 1145 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1146 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_div_4_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1147 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1148 | for (size_t k = 8; k <= 40; k += 4) { |
| 1149 | GemmMicrokernelTester() |
| 1150 | .mr(8) |
| 1151 | .nr(8) |
| 1152 | .kr(1) |
| 1153 | .sr(1) |
| 1154 | .m(8) |
| 1155 | .n(8) |
| 1156 | .k(k) |
| 1157 | .a_stride(43) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1158 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1159 | } |
| 1160 | } |
| 1161 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1162 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, k_div_4_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1163 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1164 | for (size_t k = 8; k <= 40; k += 4) { |
| 1165 | for (uint32_t m = 1; m <= 8; m++) { |
| 1166 | for (uint32_t n = 1; n <= 8; n++) { |
| 1167 | GemmMicrokernelTester() |
| 1168 | .mr(8) |
| 1169 | .nr(8) |
| 1170 | .kr(1) |
| 1171 | .sr(1) |
| 1172 | .m(m) |
| 1173 | .n(n) |
| 1174 | .k(k) |
| 1175 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1176 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1177 | } |
| 1178 | } |
| 1179 | } |
| 1180 | } |
| 1181 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1182 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1183 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1184 | for (uint32_t n = 9; n < 16; n++) { |
| 1185 | for (size_t k = 1; k <= 20; k += 5) { |
| 1186 | GemmMicrokernelTester() |
| 1187 | .mr(8) |
| 1188 | .nr(8) |
| 1189 | .kr(1) |
| 1190 | .sr(1) |
| 1191 | .m(8) |
| 1192 | .n(8) |
| 1193 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1194 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1195 | } |
| 1196 | } |
| 1197 | } |
| 1198 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1199 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8_strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1200 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1201 | for (uint32_t n = 9; n < 16; n++) { |
| 1202 | for (size_t k = 1; k <= 20; k += 5) { |
| 1203 | GemmMicrokernelTester() |
| 1204 | .mr(8) |
| 1205 | .nr(8) |
| 1206 | .kr(1) |
| 1207 | .sr(1) |
| 1208 | .m(8) |
| 1209 | .n(8) |
| 1210 | .k(k) |
| 1211 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1212 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1213 | } |
| 1214 | } |
| 1215 | } |
| 1216 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1217 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1218 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1219 | for (uint32_t n = 9; n < 16; n++) { |
| 1220 | for (size_t k = 1; k <= 20; k += 5) { |
| 1221 | GemmMicrokernelTester() |
| 1222 | .mr(8) |
| 1223 | .nr(8) |
| 1224 | .kr(1) |
| 1225 | .sr(1) |
| 1226 | .m(8) |
| 1227 | .n(n) |
| 1228 | .k(k) |
| 1229 | .a_stride(23) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1230 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1231 | } |
| 1232 | } |
| 1233 | } |
| 1234 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1235 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_gt_8_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1236 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1237 | for (uint32_t n = 9; n < 16; n++) { |
| 1238 | for (size_t k = 1; k <= 20; k += 5) { |
| 1239 | for (uint32_t m = 1; m <= 8; m++) { |
| 1240 | GemmMicrokernelTester() |
| 1241 | .mr(8) |
| 1242 | .nr(8) |
| 1243 | .kr(1) |
| 1244 | .sr(1) |
| 1245 | .m(m) |
| 1246 | .n(n) |
| 1247 | .k(k) |
| 1248 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1249 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1250 | } |
| 1251 | } |
| 1252 | } |
| 1253 | } |
| 1254 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1255 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1256 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1257 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1258 | for (size_t k = 1; k <= 20; k += 5) { |
| 1259 | GemmMicrokernelTester() |
| 1260 | .mr(8) |
| 1261 | .nr(8) |
| 1262 | .kr(1) |
| 1263 | .sr(1) |
| 1264 | .m(8) |
| 1265 | .n(8) |
| 1266 | .k(k) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1267 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1268 | } |
| 1269 | } |
| 1270 | } |
| 1271 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1272 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8_strided_cn) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1273 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1274 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1275 | for (size_t k = 1; k <= 20; k += 5) { |
| 1276 | GemmMicrokernelTester() |
| 1277 | .mr(8) |
| 1278 | .nr(8) |
| 1279 | .kr(1) |
| 1280 | .sr(1) |
| 1281 | .m(8) |
| 1282 | .n(n) |
| 1283 | .k(k) |
| 1284 | .cn_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1285 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1286 | } |
| 1287 | } |
| 1288 | } |
| 1289 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1290 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8_strided_a) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1291 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1292 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1293 | for (size_t k = 1; k <= 20; k += 5) { |
| 1294 | GemmMicrokernelTester() |
| 1295 | .mr(8) |
| 1296 | .nr(8) |
| 1297 | .kr(1) |
| 1298 | .sr(1) |
| 1299 | .m(8) |
| 1300 | .n(n) |
| 1301 | .k(k) |
| 1302 | .a_stride(23) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1303 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1304 | } |
| 1305 | } |
| 1306 | } |
| 1307 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1308 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, n_div_8_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1309 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1310 | for (uint32_t n = 16; n <= 24; n += 8) { |
| 1311 | for (size_t k = 1; k <= 20; k += 5) { |
| 1312 | for (uint32_t m = 1; m <= 8; m++) { |
| 1313 | GemmMicrokernelTester() |
| 1314 | .mr(8) |
| 1315 | .nr(8) |
| 1316 | .kr(1) |
| 1317 | .sr(1) |
| 1318 | .m(m) |
| 1319 | .n(n) |
| 1320 | .k(k) |
| 1321 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1322 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1323 | } |
| 1324 | } |
| 1325 | } |
| 1326 | } |
| 1327 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1328 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, strided_cm_subtile) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1329 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1330 | for (size_t k = 1; k <= 20; k += 5) { |
| 1331 | for (uint32_t m = 1; m <= 8; m++) { |
| 1332 | for (uint32_t n = 1; n <= 8; n++) { |
| 1333 | GemmMicrokernelTester() |
| 1334 | .mr(8) |
| 1335 | .nr(8) |
| 1336 | .kr(1) |
| 1337 | .sr(1) |
| 1338 | .m(m) |
| 1339 | .n(n) |
| 1340 | .k(k) |
| 1341 | .cm_stride(11) |
| 1342 | .iterations(1) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1343 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1344 | } |
| 1345 | } |
| 1346 | } |
| 1347 | } |
| 1348 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1349 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, qmin) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1350 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1351 | GemmMicrokernelTester() |
| 1352 | .mr(8) |
| 1353 | .nr(8) |
| 1354 | .kr(1) |
| 1355 | .sr(1) |
| 1356 | .m(8) |
| 1357 | .n(8) |
| 1358 | .k(4) |
| 1359 | .qmin(128) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1360 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1361 | } |
| 1362 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1363 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, qmax) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1364 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1365 | GemmMicrokernelTester() |
| 1366 | .mr(8) |
| 1367 | .nr(8) |
| 1368 | .kr(1) |
| 1369 | .sr(1) |
| 1370 | .m(8) |
| 1371 | .n(8) |
| 1372 | .k(4) |
| 1373 | .qmax(128) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1374 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1375 | } |
| 1376 | |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1377 | TEST(F16_GEMM_MINMAX_8X8__NEONFP16ARITH_LD64, strided_cm) { |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1378 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1379 | GemmMicrokernelTester() |
| 1380 | .mr(8) |
| 1381 | .nr(8) |
| 1382 | .kr(1) |
| 1383 | .sr(1) |
| 1384 | .m(8) |
| 1385 | .n(8) |
| 1386 | .k(4) |
| 1387 | .cm_stride(11) |
Marat Dukhan | de06f49 | 2020-04-09 00:19:31 -0700 | [diff] [blame] | 1388 | .Test(xnn_f16_gemm_minmax_ukernel_8x8__neonfp16arith_ld64); |
XNNPACK Team | b455b12 | 2019-09-27 18:10:33 -0700 | [diff] [blame] | 1389 | } |
Marat Dukhan | 1dadbf7 | 2019-10-01 10:46:20 -0700 | [diff] [blame] | 1390 | #endif // XNN_ARCH_ARM64 |
Frank Barchard | 683f559 | 2020-04-10 00:48:26 -0700 | [diff] [blame] | 1391 | |
| 1392 | |
| 1393 | #if XNN_ARCH_ARM64 |
| 1394 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2) { |
| 1395 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1396 | GemmMicrokernelTester() |
| 1397 | .mr(4) |
| 1398 | .nr(16) |
| 1399 | .kr(1) |
| 1400 | .sr(1) |
| 1401 | .m(4) |
| 1402 | .n(16) |
| 1403 | .k(2) |
| 1404 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1405 | } |
| 1406 | |
| 1407 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, strided_cn) { |
| 1408 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1409 | GemmMicrokernelTester() |
| 1410 | .mr(4) |
| 1411 | .nr(16) |
| 1412 | .kr(1) |
| 1413 | .sr(1) |
| 1414 | .m(4) |
| 1415 | .n(16) |
| 1416 | .k(2) |
| 1417 | .cn_stride(19) |
| 1418 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1419 | } |
| 1420 | |
| 1421 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_strided_a) { |
| 1422 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1423 | GemmMicrokernelTester() |
| 1424 | .mr(4) |
| 1425 | .nr(16) |
| 1426 | .kr(1) |
| 1427 | .sr(1) |
| 1428 | .m(4) |
| 1429 | .n(16) |
| 1430 | .k(2) |
| 1431 | .a_stride(5) |
| 1432 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1433 | } |
| 1434 | |
| 1435 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_subtile) { |
| 1436 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1437 | for (uint32_t m = 1; m <= 4; m++) { |
| 1438 | for (uint32_t n = 1; n <= 16; n++) { |
| 1439 | GemmMicrokernelTester() |
| 1440 | .mr(4) |
| 1441 | .nr(16) |
| 1442 | .kr(1) |
| 1443 | .sr(1) |
| 1444 | .m(m) |
| 1445 | .n(n) |
| 1446 | .k(2) |
| 1447 | .iterations(1) |
| 1448 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1449 | } |
| 1450 | } |
| 1451 | } |
| 1452 | |
| 1453 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_subtile_m) { |
| 1454 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1455 | for (uint32_t m = 1; m <= 4; m++) { |
| 1456 | GemmMicrokernelTester() |
| 1457 | .mr(4) |
| 1458 | .nr(16) |
| 1459 | .kr(1) |
| 1460 | .sr(1) |
| 1461 | .m(m) |
| 1462 | .n(16) |
| 1463 | .k(2) |
| 1464 | .iterations(1) |
| 1465 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1466 | } |
| 1467 | } |
| 1468 | |
| 1469 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_subtile_n) { |
| 1470 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1471 | for (uint32_t n = 1; n <= 16; n++) { |
| 1472 | GemmMicrokernelTester() |
| 1473 | .mr(4) |
| 1474 | .nr(16) |
| 1475 | .kr(1) |
| 1476 | .sr(1) |
| 1477 | .m(4) |
| 1478 | .n(n) |
| 1479 | .k(2) |
| 1480 | .iterations(1) |
| 1481 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1482 | } |
| 1483 | } |
| 1484 | |
| 1485 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_lt_2) { |
| 1486 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1487 | for (size_t k = 1; k < 2; k++) { |
| 1488 | GemmMicrokernelTester() |
| 1489 | .mr(4) |
| 1490 | .nr(16) |
| 1491 | .kr(1) |
| 1492 | .sr(1) |
| 1493 | .m(4) |
| 1494 | .n(16) |
| 1495 | .k(k) |
| 1496 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1497 | } |
| 1498 | } |
| 1499 | |
| 1500 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_lt_2_strided_a) { |
| 1501 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1502 | for (size_t k = 1; k < 2; k++) { |
| 1503 | GemmMicrokernelTester() |
| 1504 | .mr(4) |
| 1505 | .nr(16) |
| 1506 | .kr(1) |
| 1507 | .sr(1) |
| 1508 | .m(4) |
| 1509 | .n(16) |
| 1510 | .k(k) |
| 1511 | .a_stride(5) |
| 1512 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1513 | } |
| 1514 | } |
| 1515 | |
| 1516 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_lt_2_subtile) { |
| 1517 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1518 | for (size_t k = 1; k < 2; k++) { |
| 1519 | for (uint32_t m = 1; m <= 4; m++) { |
| 1520 | for (uint32_t n = 1; n <= 16; n++) { |
| 1521 | GemmMicrokernelTester() |
| 1522 | .mr(4) |
| 1523 | .nr(16) |
| 1524 | .kr(1) |
| 1525 | .sr(1) |
| 1526 | .m(m) |
| 1527 | .n(n) |
| 1528 | .k(k) |
| 1529 | .iterations(1) |
| 1530 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1531 | } |
| 1532 | } |
| 1533 | } |
| 1534 | } |
| 1535 | |
| 1536 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_gt_2) { |
| 1537 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1538 | for (size_t k = 3; k < 4; k++) { |
| 1539 | GemmMicrokernelTester() |
| 1540 | .mr(4) |
| 1541 | .nr(16) |
| 1542 | .kr(1) |
| 1543 | .sr(1) |
| 1544 | .m(4) |
| 1545 | .n(16) |
| 1546 | .k(k) |
| 1547 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1548 | } |
| 1549 | } |
| 1550 | |
| 1551 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_gt_2_strided_a) { |
| 1552 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1553 | for (size_t k = 3; k < 4; k++) { |
| 1554 | GemmMicrokernelTester() |
| 1555 | .mr(4) |
| 1556 | .nr(16) |
| 1557 | .kr(1) |
| 1558 | .sr(1) |
| 1559 | .m(4) |
| 1560 | .n(16) |
| 1561 | .k(k) |
| 1562 | .a_stride(7) |
| 1563 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1564 | } |
| 1565 | } |
| 1566 | |
| 1567 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_gt_2_subtile) { |
| 1568 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1569 | for (size_t k = 3; k < 4; k++) { |
| 1570 | for (uint32_t m = 1; m <= 4; m++) { |
| 1571 | for (uint32_t n = 1; n <= 16; n++) { |
| 1572 | GemmMicrokernelTester() |
| 1573 | .mr(4) |
| 1574 | .nr(16) |
| 1575 | .kr(1) |
| 1576 | .sr(1) |
| 1577 | .m(m) |
| 1578 | .n(n) |
| 1579 | .k(k) |
| 1580 | .iterations(1) |
| 1581 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1582 | } |
| 1583 | } |
| 1584 | } |
| 1585 | } |
| 1586 | |
| 1587 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_div_2) { |
| 1588 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1589 | for (size_t k = 4; k <= 20; k += 2) { |
| 1590 | GemmMicrokernelTester() |
| 1591 | .mr(4) |
| 1592 | .nr(16) |
| 1593 | .kr(1) |
| 1594 | .sr(1) |
| 1595 | .m(4) |
| 1596 | .n(16) |
| 1597 | .k(k) |
| 1598 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1599 | } |
| 1600 | } |
| 1601 | |
| 1602 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_div_2_strided_a) { |
| 1603 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1604 | for (size_t k = 4; k <= 20; k += 2) { |
| 1605 | GemmMicrokernelTester() |
| 1606 | .mr(4) |
| 1607 | .nr(16) |
| 1608 | .kr(1) |
| 1609 | .sr(1) |
| 1610 | .m(4) |
| 1611 | .n(16) |
| 1612 | .k(k) |
| 1613 | .a_stride(23) |
| 1614 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1615 | } |
| 1616 | } |
| 1617 | |
| 1618 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, k_div_2_subtile) { |
| 1619 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1620 | for (size_t k = 4; k <= 20; k += 2) { |
| 1621 | for (uint32_t m = 1; m <= 4; m++) { |
| 1622 | for (uint32_t n = 1; n <= 16; n++) { |
| 1623 | GemmMicrokernelTester() |
| 1624 | .mr(4) |
| 1625 | .nr(16) |
| 1626 | .kr(1) |
| 1627 | .sr(1) |
| 1628 | .m(m) |
| 1629 | .n(n) |
| 1630 | .k(k) |
| 1631 | .iterations(1) |
| 1632 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1633 | } |
| 1634 | } |
| 1635 | } |
| 1636 | } |
| 1637 | |
| 1638 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16) { |
| 1639 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1640 | for (uint32_t n = 17; n < 32; n++) { |
| 1641 | for (size_t k = 1; k <= 10; k += 3) { |
| 1642 | GemmMicrokernelTester() |
| 1643 | .mr(4) |
| 1644 | .nr(16) |
| 1645 | .kr(1) |
| 1646 | .sr(1) |
| 1647 | .m(4) |
| 1648 | .n(16) |
| 1649 | .k(k) |
| 1650 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1651 | } |
| 1652 | } |
| 1653 | } |
| 1654 | |
| 1655 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16_strided_cn) { |
| 1656 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1657 | for (uint32_t n = 17; n < 32; n++) { |
| 1658 | for (size_t k = 1; k <= 10; k += 3) { |
| 1659 | GemmMicrokernelTester() |
| 1660 | .mr(4) |
| 1661 | .nr(16) |
| 1662 | .kr(1) |
| 1663 | .sr(1) |
| 1664 | .m(4) |
| 1665 | .n(16) |
| 1666 | .k(k) |
| 1667 | .cn_stride(19) |
| 1668 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1669 | } |
| 1670 | } |
| 1671 | } |
| 1672 | |
| 1673 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16_strided_a) { |
| 1674 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1675 | for (uint32_t n = 17; n < 32; n++) { |
| 1676 | for (size_t k = 1; k <= 10; k += 3) { |
| 1677 | GemmMicrokernelTester() |
| 1678 | .mr(4) |
| 1679 | .nr(16) |
| 1680 | .kr(1) |
| 1681 | .sr(1) |
| 1682 | .m(4) |
| 1683 | .n(n) |
| 1684 | .k(k) |
| 1685 | .a_stride(13) |
| 1686 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1687 | } |
| 1688 | } |
| 1689 | } |
| 1690 | |
| 1691 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16_subtile) { |
| 1692 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1693 | for (uint32_t n = 17; n < 32; n++) { |
| 1694 | for (size_t k = 1; k <= 10; k += 3) { |
| 1695 | for (uint32_t m = 1; m <= 4; m++) { |
| 1696 | GemmMicrokernelTester() |
| 1697 | .mr(4) |
| 1698 | .nr(16) |
| 1699 | .kr(1) |
| 1700 | .sr(1) |
| 1701 | .m(m) |
| 1702 | .n(n) |
| 1703 | .k(k) |
| 1704 | .iterations(1) |
| 1705 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1706 | } |
| 1707 | } |
| 1708 | } |
| 1709 | } |
| 1710 | |
| 1711 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_div_16) { |
| 1712 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1713 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 1714 | for (size_t k = 1; k <= 10; k += 3) { |
| 1715 | GemmMicrokernelTester() |
| 1716 | .mr(4) |
| 1717 | .nr(16) |
| 1718 | .kr(1) |
| 1719 | .sr(1) |
| 1720 | .m(4) |
| 1721 | .n(16) |
| 1722 | .k(k) |
| 1723 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1724 | } |
| 1725 | } |
| 1726 | } |
| 1727 | |
| 1728 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_div_16_strided_cn) { |
| 1729 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1730 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 1731 | for (size_t k = 1; k <= 10; k += 3) { |
| 1732 | GemmMicrokernelTester() |
| 1733 | .mr(4) |
| 1734 | .nr(16) |
| 1735 | .kr(1) |
| 1736 | .sr(1) |
| 1737 | .m(4) |
| 1738 | .n(n) |
| 1739 | .k(k) |
| 1740 | .cn_stride(19) |
| 1741 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1742 | } |
| 1743 | } |
| 1744 | } |
| 1745 | |
| 1746 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_div_16_strided_a) { |
| 1747 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1748 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 1749 | for (size_t k = 1; k <= 10; k += 3) { |
| 1750 | GemmMicrokernelTester() |
| 1751 | .mr(4) |
| 1752 | .nr(16) |
| 1753 | .kr(1) |
| 1754 | .sr(1) |
| 1755 | .m(4) |
| 1756 | .n(n) |
| 1757 | .k(k) |
| 1758 | .a_stride(13) |
| 1759 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1760 | } |
| 1761 | } |
| 1762 | } |
| 1763 | |
| 1764 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, n_div_16_subtile) { |
| 1765 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1766 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 1767 | for (size_t k = 1; k <= 10; k += 3) { |
| 1768 | for (uint32_t m = 1; m <= 4; m++) { |
| 1769 | GemmMicrokernelTester() |
| 1770 | .mr(4) |
| 1771 | .nr(16) |
| 1772 | .kr(1) |
| 1773 | .sr(1) |
| 1774 | .m(m) |
| 1775 | .n(n) |
| 1776 | .k(k) |
| 1777 | .iterations(1) |
| 1778 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1779 | } |
| 1780 | } |
| 1781 | } |
| 1782 | } |
| 1783 | |
| 1784 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, strided_cm_subtile) { |
| 1785 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1786 | for (size_t k = 1; k <= 10; k += 3) { |
| 1787 | for (uint32_t m = 1; m <= 4; m++) { |
| 1788 | for (uint32_t n = 1; n <= 16; n++) { |
| 1789 | GemmMicrokernelTester() |
| 1790 | .mr(4) |
| 1791 | .nr(16) |
| 1792 | .kr(1) |
| 1793 | .sr(1) |
| 1794 | .m(m) |
| 1795 | .n(n) |
| 1796 | .k(k) |
| 1797 | .cm_stride(19) |
| 1798 | .iterations(1) |
| 1799 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1800 | } |
| 1801 | } |
| 1802 | } |
| 1803 | } |
| 1804 | |
| 1805 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, qmin) { |
| 1806 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1807 | GemmMicrokernelTester() |
| 1808 | .mr(4) |
| 1809 | .nr(16) |
| 1810 | .kr(1) |
| 1811 | .sr(1) |
| 1812 | .m(4) |
| 1813 | .n(16) |
| 1814 | .k(2) |
| 1815 | .qmin(128) |
| 1816 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1817 | } |
| 1818 | |
| 1819 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, qmax) { |
| 1820 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1821 | GemmMicrokernelTester() |
| 1822 | .mr(4) |
| 1823 | .nr(16) |
| 1824 | .kr(1) |
| 1825 | .sr(1) |
| 1826 | .m(4) |
| 1827 | .n(16) |
| 1828 | .k(2) |
| 1829 | .qmax(128) |
| 1830 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1831 | } |
| 1832 | |
| 1833 | TEST(F16_GEMM_MINMAX_4X16__AARCH64_NEONFP16ARITH_LD32, strided_cm) { |
| 1834 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1835 | GemmMicrokernelTester() |
| 1836 | .mr(4) |
| 1837 | .nr(16) |
| 1838 | .kr(1) |
| 1839 | .sr(1) |
| 1840 | .m(4) |
| 1841 | .n(16) |
| 1842 | .k(2) |
| 1843 | .cm_stride(19) |
| 1844 | .Test(xnn_f16_gemm_minmax_ukernel_4x16__aarch64_neonfp16arith_ld32); |
| 1845 | } |
| 1846 | #endif // XNN_ARCH_ARM64 |
| 1847 | |
| 1848 | |
| 1849 | #if XNN_ARCH_ARM64 |
| 1850 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2) { |
| 1851 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1852 | GemmMicrokernelTester() |
| 1853 | .mr(6) |
| 1854 | .nr(16) |
| 1855 | .kr(1) |
| 1856 | .sr(1) |
| 1857 | .m(6) |
| 1858 | .n(16) |
| 1859 | .k(2) |
| 1860 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1861 | } |
| 1862 | |
| 1863 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, strided_cn) { |
| 1864 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1865 | GemmMicrokernelTester() |
| 1866 | .mr(6) |
| 1867 | .nr(16) |
| 1868 | .kr(1) |
| 1869 | .sr(1) |
| 1870 | .m(6) |
| 1871 | .n(16) |
| 1872 | .k(2) |
| 1873 | .cn_stride(19) |
| 1874 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1875 | } |
| 1876 | |
| 1877 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_strided_a) { |
| 1878 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1879 | GemmMicrokernelTester() |
| 1880 | .mr(6) |
| 1881 | .nr(16) |
| 1882 | .kr(1) |
| 1883 | .sr(1) |
| 1884 | .m(6) |
| 1885 | .n(16) |
| 1886 | .k(2) |
| 1887 | .a_stride(5) |
| 1888 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1889 | } |
| 1890 | |
| 1891 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_subtile) { |
| 1892 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1893 | for (uint32_t m = 1; m <= 6; m++) { |
| 1894 | for (uint32_t n = 1; n <= 16; n++) { |
| 1895 | GemmMicrokernelTester() |
| 1896 | .mr(6) |
| 1897 | .nr(16) |
| 1898 | .kr(1) |
| 1899 | .sr(1) |
| 1900 | .m(m) |
| 1901 | .n(n) |
| 1902 | .k(2) |
| 1903 | .iterations(1) |
| 1904 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1905 | } |
| 1906 | } |
| 1907 | } |
| 1908 | |
| 1909 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_subtile_m) { |
| 1910 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1911 | for (uint32_t m = 1; m <= 6; m++) { |
| 1912 | GemmMicrokernelTester() |
| 1913 | .mr(6) |
| 1914 | .nr(16) |
| 1915 | .kr(1) |
| 1916 | .sr(1) |
| 1917 | .m(m) |
| 1918 | .n(16) |
| 1919 | .k(2) |
| 1920 | .iterations(1) |
| 1921 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1922 | } |
| 1923 | } |
| 1924 | |
| 1925 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_eq_2_subtile_n) { |
| 1926 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1927 | for (uint32_t n = 1; n <= 16; n++) { |
| 1928 | GemmMicrokernelTester() |
| 1929 | .mr(6) |
| 1930 | .nr(16) |
| 1931 | .kr(1) |
| 1932 | .sr(1) |
| 1933 | .m(6) |
| 1934 | .n(n) |
| 1935 | .k(2) |
| 1936 | .iterations(1) |
| 1937 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1938 | } |
| 1939 | } |
| 1940 | |
| 1941 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_lt_2) { |
| 1942 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1943 | for (size_t k = 1; k < 2; k++) { |
| 1944 | GemmMicrokernelTester() |
| 1945 | .mr(6) |
| 1946 | .nr(16) |
| 1947 | .kr(1) |
| 1948 | .sr(1) |
| 1949 | .m(6) |
| 1950 | .n(16) |
| 1951 | .k(k) |
| 1952 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1953 | } |
| 1954 | } |
| 1955 | |
| 1956 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_lt_2_strided_a) { |
| 1957 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1958 | for (size_t k = 1; k < 2; k++) { |
| 1959 | GemmMicrokernelTester() |
| 1960 | .mr(6) |
| 1961 | .nr(16) |
| 1962 | .kr(1) |
| 1963 | .sr(1) |
| 1964 | .m(6) |
| 1965 | .n(16) |
| 1966 | .k(k) |
| 1967 | .a_stride(5) |
| 1968 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1969 | } |
| 1970 | } |
| 1971 | |
| 1972 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_lt_2_subtile) { |
| 1973 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1974 | for (size_t k = 1; k < 2; k++) { |
| 1975 | for (uint32_t m = 1; m <= 6; m++) { |
| 1976 | for (uint32_t n = 1; n <= 16; n++) { |
| 1977 | GemmMicrokernelTester() |
| 1978 | .mr(6) |
| 1979 | .nr(16) |
| 1980 | .kr(1) |
| 1981 | .sr(1) |
| 1982 | .m(m) |
| 1983 | .n(n) |
| 1984 | .k(k) |
| 1985 | .iterations(1) |
| 1986 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 1987 | } |
| 1988 | } |
| 1989 | } |
| 1990 | } |
| 1991 | |
| 1992 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_gt_2) { |
| 1993 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 1994 | for (size_t k = 3; k < 4; k++) { |
| 1995 | GemmMicrokernelTester() |
| 1996 | .mr(6) |
| 1997 | .nr(16) |
| 1998 | .kr(1) |
| 1999 | .sr(1) |
| 2000 | .m(6) |
| 2001 | .n(16) |
| 2002 | .k(k) |
| 2003 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2004 | } |
| 2005 | } |
| 2006 | |
| 2007 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_gt_2_strided_a) { |
| 2008 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2009 | for (size_t k = 3; k < 4; k++) { |
| 2010 | GemmMicrokernelTester() |
| 2011 | .mr(6) |
| 2012 | .nr(16) |
| 2013 | .kr(1) |
| 2014 | .sr(1) |
| 2015 | .m(6) |
| 2016 | .n(16) |
| 2017 | .k(k) |
| 2018 | .a_stride(7) |
| 2019 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2020 | } |
| 2021 | } |
| 2022 | |
| 2023 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_gt_2_subtile) { |
| 2024 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2025 | for (size_t k = 3; k < 4; k++) { |
| 2026 | for (uint32_t m = 1; m <= 6; m++) { |
| 2027 | for (uint32_t n = 1; n <= 16; n++) { |
| 2028 | GemmMicrokernelTester() |
| 2029 | .mr(6) |
| 2030 | .nr(16) |
| 2031 | .kr(1) |
| 2032 | .sr(1) |
| 2033 | .m(m) |
| 2034 | .n(n) |
| 2035 | .k(k) |
| 2036 | .iterations(1) |
| 2037 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2038 | } |
| 2039 | } |
| 2040 | } |
| 2041 | } |
| 2042 | |
| 2043 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_div_2) { |
| 2044 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2045 | for (size_t k = 4; k <= 20; k += 2) { |
| 2046 | GemmMicrokernelTester() |
| 2047 | .mr(6) |
| 2048 | .nr(16) |
| 2049 | .kr(1) |
| 2050 | .sr(1) |
| 2051 | .m(6) |
| 2052 | .n(16) |
| 2053 | .k(k) |
| 2054 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2055 | } |
| 2056 | } |
| 2057 | |
| 2058 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_div_2_strided_a) { |
| 2059 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2060 | for (size_t k = 4; k <= 20; k += 2) { |
| 2061 | GemmMicrokernelTester() |
| 2062 | .mr(6) |
| 2063 | .nr(16) |
| 2064 | .kr(1) |
| 2065 | .sr(1) |
| 2066 | .m(6) |
| 2067 | .n(16) |
| 2068 | .k(k) |
| 2069 | .a_stride(23) |
| 2070 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2071 | } |
| 2072 | } |
| 2073 | |
| 2074 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, k_div_2_subtile) { |
| 2075 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2076 | for (size_t k = 4; k <= 20; k += 2) { |
| 2077 | for (uint32_t m = 1; m <= 6; m++) { |
| 2078 | for (uint32_t n = 1; n <= 16; n++) { |
| 2079 | GemmMicrokernelTester() |
| 2080 | .mr(6) |
| 2081 | .nr(16) |
| 2082 | .kr(1) |
| 2083 | .sr(1) |
| 2084 | .m(m) |
| 2085 | .n(n) |
| 2086 | .k(k) |
| 2087 | .iterations(1) |
| 2088 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2089 | } |
| 2090 | } |
| 2091 | } |
| 2092 | } |
| 2093 | |
| 2094 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16) { |
| 2095 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2096 | for (uint32_t n = 17; n < 32; n++) { |
| 2097 | for (size_t k = 1; k <= 10; k += 3) { |
| 2098 | GemmMicrokernelTester() |
| 2099 | .mr(6) |
| 2100 | .nr(16) |
| 2101 | .kr(1) |
| 2102 | .sr(1) |
| 2103 | .m(6) |
| 2104 | .n(16) |
| 2105 | .k(k) |
| 2106 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2107 | } |
| 2108 | } |
| 2109 | } |
| 2110 | |
| 2111 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16_strided_cn) { |
| 2112 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2113 | for (uint32_t n = 17; n < 32; n++) { |
| 2114 | for (size_t k = 1; k <= 10; k += 3) { |
| 2115 | GemmMicrokernelTester() |
| 2116 | .mr(6) |
| 2117 | .nr(16) |
| 2118 | .kr(1) |
| 2119 | .sr(1) |
| 2120 | .m(6) |
| 2121 | .n(16) |
| 2122 | .k(k) |
| 2123 | .cn_stride(19) |
| 2124 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2125 | } |
| 2126 | } |
| 2127 | } |
| 2128 | |
| 2129 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16_strided_a) { |
| 2130 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2131 | for (uint32_t n = 17; n < 32; n++) { |
| 2132 | for (size_t k = 1; k <= 10; k += 3) { |
| 2133 | GemmMicrokernelTester() |
| 2134 | .mr(6) |
| 2135 | .nr(16) |
| 2136 | .kr(1) |
| 2137 | .sr(1) |
| 2138 | .m(6) |
| 2139 | .n(n) |
| 2140 | .k(k) |
| 2141 | .a_stride(13) |
| 2142 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2143 | } |
| 2144 | } |
| 2145 | } |
| 2146 | |
| 2147 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_gt_16_subtile) { |
| 2148 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2149 | for (uint32_t n = 17; n < 32; n++) { |
| 2150 | for (size_t k = 1; k <= 10; k += 3) { |
| 2151 | for (uint32_t m = 1; m <= 6; m++) { |
| 2152 | GemmMicrokernelTester() |
| 2153 | .mr(6) |
| 2154 | .nr(16) |
| 2155 | .kr(1) |
| 2156 | .sr(1) |
| 2157 | .m(m) |
| 2158 | .n(n) |
| 2159 | .k(k) |
| 2160 | .iterations(1) |
| 2161 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2162 | } |
| 2163 | } |
| 2164 | } |
| 2165 | } |
| 2166 | |
| 2167 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_div_16) { |
| 2168 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2169 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2170 | for (size_t k = 1; k <= 10; k += 3) { |
| 2171 | GemmMicrokernelTester() |
| 2172 | .mr(6) |
| 2173 | .nr(16) |
| 2174 | .kr(1) |
| 2175 | .sr(1) |
| 2176 | .m(6) |
| 2177 | .n(16) |
| 2178 | .k(k) |
| 2179 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2180 | } |
| 2181 | } |
| 2182 | } |
| 2183 | |
| 2184 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_div_16_strided_cn) { |
| 2185 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2186 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2187 | for (size_t k = 1; k <= 10; k += 3) { |
| 2188 | GemmMicrokernelTester() |
| 2189 | .mr(6) |
| 2190 | .nr(16) |
| 2191 | .kr(1) |
| 2192 | .sr(1) |
| 2193 | .m(6) |
| 2194 | .n(n) |
| 2195 | .k(k) |
| 2196 | .cn_stride(19) |
| 2197 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2198 | } |
| 2199 | } |
| 2200 | } |
| 2201 | |
| 2202 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_div_16_strided_a) { |
| 2203 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2204 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2205 | for (size_t k = 1; k <= 10; k += 3) { |
| 2206 | GemmMicrokernelTester() |
| 2207 | .mr(6) |
| 2208 | .nr(16) |
| 2209 | .kr(1) |
| 2210 | .sr(1) |
| 2211 | .m(6) |
| 2212 | .n(n) |
| 2213 | .k(k) |
| 2214 | .a_stride(13) |
| 2215 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2216 | } |
| 2217 | } |
| 2218 | } |
| 2219 | |
| 2220 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, n_div_16_subtile) { |
| 2221 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2222 | for (uint32_t n = 32; n <= 48; n += 16) { |
| 2223 | for (size_t k = 1; k <= 10; k += 3) { |
| 2224 | for (uint32_t m = 1; m <= 6; m++) { |
| 2225 | GemmMicrokernelTester() |
| 2226 | .mr(6) |
| 2227 | .nr(16) |
| 2228 | .kr(1) |
| 2229 | .sr(1) |
| 2230 | .m(m) |
| 2231 | .n(n) |
| 2232 | .k(k) |
| 2233 | .iterations(1) |
| 2234 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2235 | } |
| 2236 | } |
| 2237 | } |
| 2238 | } |
| 2239 | |
| 2240 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, strided_cm_subtile) { |
| 2241 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2242 | for (size_t k = 1; k <= 10; k += 3) { |
| 2243 | for (uint32_t m = 1; m <= 6; m++) { |
| 2244 | for (uint32_t n = 1; n <= 16; n++) { |
| 2245 | GemmMicrokernelTester() |
| 2246 | .mr(6) |
| 2247 | .nr(16) |
| 2248 | .kr(1) |
| 2249 | .sr(1) |
| 2250 | .m(m) |
| 2251 | .n(n) |
| 2252 | .k(k) |
| 2253 | .cm_stride(19) |
| 2254 | .iterations(1) |
| 2255 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2256 | } |
| 2257 | } |
| 2258 | } |
| 2259 | } |
| 2260 | |
| 2261 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, qmin) { |
| 2262 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2263 | GemmMicrokernelTester() |
| 2264 | .mr(6) |
| 2265 | .nr(16) |
| 2266 | .kr(1) |
| 2267 | .sr(1) |
| 2268 | .m(6) |
| 2269 | .n(16) |
| 2270 | .k(2) |
| 2271 | .qmin(128) |
| 2272 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2273 | } |
| 2274 | |
| 2275 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, qmax) { |
| 2276 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2277 | GemmMicrokernelTester() |
| 2278 | .mr(6) |
| 2279 | .nr(16) |
| 2280 | .kr(1) |
| 2281 | .sr(1) |
| 2282 | .m(6) |
| 2283 | .n(16) |
| 2284 | .k(2) |
| 2285 | .qmax(128) |
| 2286 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2287 | } |
| 2288 | |
| 2289 | TEST(F16_GEMM_MINMAX_6X16__AARCH64_NEONFP16ARITH_LD32, strided_cm) { |
| 2290 | TEST_REQUIRES_ARM_NEON_FP16_ARITH; |
| 2291 | GemmMicrokernelTester() |
| 2292 | .mr(6) |
| 2293 | .nr(16) |
| 2294 | .kr(1) |
| 2295 | .sr(1) |
| 2296 | .m(6) |
| 2297 | .n(16) |
| 2298 | .k(2) |
| 2299 | .cm_stride(19) |
| 2300 | .Test(xnn_f16_gemm_minmax_ukernel_6x16__aarch64_neonfp16arith_ld32); |
| 2301 | } |
| 2302 | #endif // XNN_ARCH_ARM64 |