blob: 43b9b3a46e3bfece4c25da93559a1549ecbfdb35 [file] [log] [blame]
Zhi An Nge96b6bc2022-02-03 10:49:46 -08001// 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-igemm-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
26#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
27 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, 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_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
38 }
39
40 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, 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_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
52 }
53
54 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8_subtile) {
55 TEST_REQUIRES_ARM_NEON;
56 for (uint32_t n = 1; n <= 8; n++) {
57 for (uint32_t m = 1; m <= 4; m++) {
58 GemmMicrokernelTester()
59 .mr(4)
60 .nr(8)
61 .kr(1)
62 .sr(1)
63 .m(m)
64 .n(n)
65 .k(8)
66 .iterations(1)
67 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
68 }
69 }
70 }
71
72 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8_subtile_m) {
73 TEST_REQUIRES_ARM_NEON;
74 for (uint32_t m = 1; m <= 4; m++) {
75 GemmMicrokernelTester()
76 .mr(4)
77 .nr(8)
78 .kr(1)
79 .sr(1)
80 .m(m)
81 .n(8)
82 .k(8)
83 .iterations(1)
84 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
85 }
86 }
87
88 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_eq_8_subtile_n) {
89 TEST_REQUIRES_ARM_NEON;
90 for (uint32_t n = 1; n <= 8; n++) {
91 GemmMicrokernelTester()
92 .mr(4)
93 .nr(8)
94 .kr(1)
95 .sr(1)
96 .m(4)
97 .n(n)
98 .k(8)
99 .iterations(1)
100 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
101 }
102 }
103
104 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_lt_8) {
105 TEST_REQUIRES_ARM_NEON;
106 for (size_t k = 1; k < 8; k++) {
107 GemmMicrokernelTester()
108 .mr(4)
109 .nr(8)
110 .kr(1)
111 .sr(1)
112 .m(4)
113 .n(8)
114 .k(k)
115 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
116 }
117 }
118
119 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_lt_8_subtile) {
120 TEST_REQUIRES_ARM_NEON;
121 for (size_t k = 1; k < 8; k++) {
122 for (uint32_t n = 1; n <= 8; n++) {
123 for (uint32_t m = 1; m <= 4; m++) {
124 GemmMicrokernelTester()
125 .mr(4)
126 .nr(8)
127 .kr(1)
128 .sr(1)
129 .m(m)
130 .n(n)
131 .k(k)
132 .iterations(1)
133 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
134 }
135 }
136 }
137 }
138
139 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_gt_8) {
140 TEST_REQUIRES_ARM_NEON;
141 for (size_t k = 9; k < 16; k++) {
142 GemmMicrokernelTester()
143 .mr(4)
144 .nr(8)
145 .kr(1)
146 .sr(1)
147 .m(4)
148 .n(8)
149 .k(k)
150 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
151 }
152 }
153
154 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_gt_8_subtile) {
155 TEST_REQUIRES_ARM_NEON;
156 for (size_t k = 9; k < 16; k++) {
157 for (uint32_t n = 1; n <= 8; n++) {
158 for (uint32_t m = 1; m <= 4; m++) {
159 GemmMicrokernelTester()
160 .mr(4)
161 .nr(8)
162 .kr(1)
163 .sr(1)
164 .m(m)
165 .n(n)
166 .k(k)
167 .iterations(1)
168 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
169 }
170 }
171 }
172 }
173
174 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_div_8) {
175 TEST_REQUIRES_ARM_NEON;
176 for (size_t k = 16; k <= 80; k += 8) {
177 GemmMicrokernelTester()
178 .mr(4)
179 .nr(8)
180 .kr(1)
181 .sr(1)
182 .m(4)
183 .n(8)
184 .k(k)
185 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
186 }
187 }
188
189 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, k_div_8_subtile) {
190 TEST_REQUIRES_ARM_NEON;
191 for (size_t k = 16; k <= 80; k += 8) {
192 for (uint32_t n = 1; n <= 8; n++) {
193 for (uint32_t m = 1; m <= 4; m++) {
194 GemmMicrokernelTester()
195 .mr(4)
196 .nr(8)
197 .kr(1)
198 .sr(1)
199 .m(m)
200 .n(n)
201 .k(k)
202 .iterations(1)
203 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
204 }
205 }
206 }
207 }
208
209 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8) {
210 TEST_REQUIRES_ARM_NEON;
211 for (uint32_t n = 9; n < 16; n++) {
212 for (size_t k = 1; k <= 40; k += 9) {
213 GemmMicrokernelTester()
214 .mr(4)
215 .nr(8)
216 .kr(1)
217 .sr(1)
218 .m(4)
219 .n(n)
220 .k(k)
221 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
222 }
223 }
224 }
225
226 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8_strided_cn) {
227 TEST_REQUIRES_ARM_NEON;
228 for (uint32_t n = 9; n < 16; n++) {
229 for (size_t k = 1; k <= 40; k += 9) {
230 GemmMicrokernelTester()
231 .mr(4)
232 .nr(8)
233 .kr(1)
234 .sr(1)
235 .m(4)
236 .n(n)
237 .k(k)
238 .cn_stride(11)
239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
240 }
241 }
242 }
243
244 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8_subtile) {
245 TEST_REQUIRES_ARM_NEON;
246 for (uint32_t n = 9; n < 16; n++) {
247 for (size_t k = 1; k <= 40; k += 9) {
248 for (uint32_t m = 1; m <= 4; m++) {
249 GemmMicrokernelTester()
250 .mr(4)
251 .nr(8)
252 .kr(1)
253 .sr(1)
254 .m(m)
255 .n(n)
256 .k(k)
257 .iterations(1)
258 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
259 }
260 }
261 }
262 }
263
264 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8) {
265 TEST_REQUIRES_ARM_NEON;
266 for (uint32_t n = 16; n <= 24; n += 8) {
267 for (size_t k = 1; k <= 40; k += 9) {
268 GemmMicrokernelTester()
269 .mr(4)
270 .nr(8)
271 .kr(1)
272 .sr(1)
273 .m(4)
274 .n(n)
275 .k(k)
276 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
277 }
278 }
279 }
280
281 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8_strided_cn) {
282 TEST_REQUIRES_ARM_NEON;
283 for (uint32_t n = 16; n <= 24; n += 8) {
284 for (size_t k = 1; k <= 40; k += 9) {
285 GemmMicrokernelTester()
286 .mr(4)
287 .nr(8)
288 .kr(1)
289 .sr(1)
290 .m(4)
291 .n(n)
292 .k(k)
293 .cn_stride(11)
294 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
295 }
296 }
297 }
298
299 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8_subtile) {
300 TEST_REQUIRES_ARM_NEON;
301 for (uint32_t n = 16; n <= 24; n += 8) {
302 for (size_t k = 1; k <= 40; k += 9) {
303 for (uint32_t m = 1; m <= 4; m++) {
304 GemmMicrokernelTester()
305 .mr(4)
306 .nr(8)
307 .kr(1)
308 .sr(1)
309 .m(m)
310 .n(n)
311 .k(k)
312 .iterations(1)
313 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
314 }
315 }
316 }
317 }
318
319 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, small_kernel) {
320 TEST_REQUIRES_ARM_NEON;
321 for (size_t k = 1; k <= 40; k += 9) {
322 GemmMicrokernelTester()
323 .mr(4)
324 .nr(8)
325 .kr(1)
326 .sr(1)
327 .m(4)
328 .n(8)
329 .k(k)
330 .ks(3)
331 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
332 }
333 }
334
335 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, small_kernel_subtile) {
336 TEST_REQUIRES_ARM_NEON;
337 for (size_t k = 1; k <= 40; k += 9) {
338 for (uint32_t n = 1; n <= 8; n++) {
339 for (uint32_t m = 1; m <= 4; m++) {
340 GemmMicrokernelTester()
341 .mr(4)
342 .nr(8)
343 .kr(1)
344 .sr(1)
345 .m(m)
346 .n(n)
347 .k(k)
348 .ks(3)
349 .iterations(1)
350 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
351 }
352 }
353 }
354 }
355
356 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_gt_8_small_kernel) {
357 TEST_REQUIRES_ARM_NEON;
358 for (uint32_t n = 9; n < 16; n++) {
359 for (size_t k = 1; k <= 40; k += 9) {
360 GemmMicrokernelTester()
361 .mr(4)
362 .nr(8)
363 .kr(1)
364 .sr(1)
365 .m(4)
366 .n(n)
367 .k(k)
368 .ks(3)
369 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
370 }
371 }
372 }
373
374 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, n_div_8_small_kernel) {
375 TEST_REQUIRES_ARM_NEON;
376 for (uint32_t n = 16; n <= 24; n += 8) {
377 for (size_t k = 1; k <= 40; k += 9) {
378 GemmMicrokernelTester()
379 .mr(4)
380 .nr(8)
381 .kr(1)
382 .sr(1)
383 .m(4)
384 .n(n)
385 .k(k)
386 .ks(3)
387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
388 }
389 }
390 }
391
392 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, strided_cm_subtile) {
393 TEST_REQUIRES_ARM_NEON;
394 for (size_t k = 1; k <= 40; k += 9) {
395 for (uint32_t n = 1; n <= 8; n++) {
396 for (uint32_t m = 1; m <= 4; m++) {
397 GemmMicrokernelTester()
398 .mr(4)
399 .nr(8)
400 .kr(1)
401 .sr(1)
402 .m(m)
403 .n(n)
404 .k(k)
405 .cm_stride(11)
406 .iterations(1)
407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
408 }
409 }
410 }
411 }
412
413 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, a_offset) {
414 TEST_REQUIRES_ARM_NEON;
415 for (size_t k = 1; k <= 40; k += 9) {
416 GemmMicrokernelTester()
417 .mr(4)
418 .nr(8)
419 .kr(1)
420 .sr(1)
421 .m(4)
422 .n(8)
423 .k(k)
424 .ks(3)
425 .a_offset(163)
426 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
427 }
428 }
429
430 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, zero) {
431 TEST_REQUIRES_ARM_NEON;
432 for (size_t k = 1; k <= 40; k += 9) {
433 for (uint32_t mz = 0; mz < 4; mz++) {
434 GemmMicrokernelTester()
435 .mr(4)
436 .nr(8)
437 .kr(1)
438 .sr(1)
439 .m(4)
440 .n(8)
441 .k(k)
442 .ks(3)
443 .a_offset(163)
444 .zero_index(mz)
445 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
446 }
447 }
448 }
449
450 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, qmin) {
451 TEST_REQUIRES_ARM_NEON;
452 GemmMicrokernelTester()
453 .mr(4)
454 .nr(8)
455 .kr(1)
456 .sr(1)
457 .m(4)
458 .n(8)
459 .k(8)
460 .qmin(128)
461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
462 }
463
464 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, qmax) {
465 TEST_REQUIRES_ARM_NEON;
466 GemmMicrokernelTester()
467 .mr(4)
468 .nr(8)
469 .kr(1)
470 .sr(1)
471 .m(4)
472 .n(8)
473 .k(8)
474 .qmax(128)
475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
476 }
477
478 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8__AARCH64_NEON_MLAL_LANE_LD64, strided_cm) {
479 TEST_REQUIRES_ARM_NEON;
480 GemmMicrokernelTester()
481 .mr(4)
482 .nr(8)
483 .kr(1)
484 .sr(1)
485 .m(4)
486 .n(8)
487 .k(8)
488 .cm_stride(11)
489 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8__aarch64_neon_mlal_lane_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
490 }
491#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
492
493
494#if XNN_ARCH_ARM || XNN_ARCH_ARM64
495 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8) {
496 TEST_REQUIRES_ARM_NEON;
497 GemmMicrokernelTester()
498 .mr(1)
499 .nr(8)
500 .kr(2)
501 .sr(1)
502 .m(1)
503 .n(8)
504 .k(8)
505 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
506 }
507
508 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, strided_cn) {
509 TEST_REQUIRES_ARM_NEON;
510 GemmMicrokernelTester()
511 .mr(1)
512 .nr(8)
513 .kr(2)
514 .sr(1)
515 .m(1)
516 .n(8)
517 .k(8)
518 .cn_stride(11)
519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
520 }
521
522 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile) {
523 TEST_REQUIRES_ARM_NEON;
524 for (uint32_t n = 1; n <= 8; n++) {
525 for (uint32_t m = 1; m <= 1; m++) {
526 GemmMicrokernelTester()
527 .mr(1)
528 .nr(8)
529 .kr(2)
530 .sr(1)
531 .m(m)
532 .n(n)
533 .k(8)
534 .iterations(1)
535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
536 }
537 }
538 }
539
540 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile_m) {
541 TEST_REQUIRES_ARM_NEON;
542 for (uint32_t m = 1; m <= 1; m++) {
543 GemmMicrokernelTester()
544 .mr(1)
545 .nr(8)
546 .kr(2)
547 .sr(1)
548 .m(m)
549 .n(8)
550 .k(8)
551 .iterations(1)
552 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
553 }
554 }
555
556 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_eq_8_subtile_n) {
557 TEST_REQUIRES_ARM_NEON;
558 for (uint32_t n = 1; n <= 8; n++) {
559 GemmMicrokernelTester()
560 .mr(1)
561 .nr(8)
562 .kr(2)
563 .sr(1)
564 .m(1)
565 .n(n)
566 .k(8)
567 .iterations(1)
568 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
569 }
570 }
571
572 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_lt_8) {
573 TEST_REQUIRES_ARM_NEON;
574 for (size_t k = 1; k < 8; k++) {
575 GemmMicrokernelTester()
576 .mr(1)
577 .nr(8)
578 .kr(2)
579 .sr(1)
580 .m(1)
581 .n(8)
582 .k(k)
583 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
584 }
585 }
586
587 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_lt_8_subtile) {
588 TEST_REQUIRES_ARM_NEON;
589 for (size_t k = 1; k < 8; k++) {
590 for (uint32_t n = 1; n <= 8; n++) {
591 for (uint32_t m = 1; m <= 1; m++) {
592 GemmMicrokernelTester()
593 .mr(1)
594 .nr(8)
595 .kr(2)
596 .sr(1)
597 .m(m)
598 .n(n)
599 .k(k)
600 .iterations(1)
601 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
602 }
603 }
604 }
605 }
606
607 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_gt_8) {
608 TEST_REQUIRES_ARM_NEON;
609 for (size_t k = 9; k < 16; k++) {
610 GemmMicrokernelTester()
611 .mr(1)
612 .nr(8)
613 .kr(2)
614 .sr(1)
615 .m(1)
616 .n(8)
617 .k(k)
618 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
619 }
620 }
621
622 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_gt_8_subtile) {
623 TEST_REQUIRES_ARM_NEON;
624 for (size_t k = 9; k < 16; k++) {
625 for (uint32_t n = 1; n <= 8; n++) {
626 for (uint32_t m = 1; m <= 1; m++) {
627 GemmMicrokernelTester()
628 .mr(1)
629 .nr(8)
630 .kr(2)
631 .sr(1)
632 .m(m)
633 .n(n)
634 .k(k)
635 .iterations(1)
636 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
637 }
638 }
639 }
640 }
641
642 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_div_8) {
643 TEST_REQUIRES_ARM_NEON;
644 for (size_t k = 16; k <= 80; k += 8) {
645 GemmMicrokernelTester()
646 .mr(1)
647 .nr(8)
648 .kr(2)
649 .sr(1)
650 .m(1)
651 .n(8)
652 .k(k)
653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
654 }
655 }
656
657 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, k_div_8_subtile) {
658 TEST_REQUIRES_ARM_NEON;
659 for (size_t k = 16; k <= 80; k += 8) {
660 for (uint32_t n = 1; n <= 8; n++) {
661 for (uint32_t m = 1; m <= 1; m++) {
662 GemmMicrokernelTester()
663 .mr(1)
664 .nr(8)
665 .kr(2)
666 .sr(1)
667 .m(m)
668 .n(n)
669 .k(k)
670 .iterations(1)
671 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
672 }
673 }
674 }
675 }
676
677 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8) {
678 TEST_REQUIRES_ARM_NEON;
679 for (uint32_t n = 9; n < 16; n++) {
680 for (size_t k = 1; k <= 40; k += 9) {
681 GemmMicrokernelTester()
682 .mr(1)
683 .nr(8)
684 .kr(2)
685 .sr(1)
686 .m(1)
687 .n(n)
688 .k(k)
689 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
690 }
691 }
692 }
693
694 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_strided_cn) {
695 TEST_REQUIRES_ARM_NEON;
696 for (uint32_t n = 9; n < 16; n++) {
697 for (size_t k = 1; k <= 40; k += 9) {
698 GemmMicrokernelTester()
699 .mr(1)
700 .nr(8)
701 .kr(2)
702 .sr(1)
703 .m(1)
704 .n(n)
705 .k(k)
706 .cn_stride(11)
707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
708 }
709 }
710 }
711
712 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_subtile) {
713 TEST_REQUIRES_ARM_NEON;
714 for (uint32_t n = 9; n < 16; n++) {
715 for (size_t k = 1; k <= 40; k += 9) {
716 for (uint32_t m = 1; m <= 1; m++) {
717 GemmMicrokernelTester()
718 .mr(1)
719 .nr(8)
720 .kr(2)
721 .sr(1)
722 .m(m)
723 .n(n)
724 .k(k)
725 .iterations(1)
726 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
727 }
728 }
729 }
730 }
731
732 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8) {
733 TEST_REQUIRES_ARM_NEON;
734 for (uint32_t n = 16; n <= 24; n += 8) {
735 for (size_t k = 1; k <= 40; k += 9) {
736 GemmMicrokernelTester()
737 .mr(1)
738 .nr(8)
739 .kr(2)
740 .sr(1)
741 .m(1)
742 .n(n)
743 .k(k)
744 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
745 }
746 }
747 }
748
749 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_strided_cn) {
750 TEST_REQUIRES_ARM_NEON;
751 for (uint32_t n = 16; n <= 24; n += 8) {
752 for (size_t k = 1; k <= 40; k += 9) {
753 GemmMicrokernelTester()
754 .mr(1)
755 .nr(8)
756 .kr(2)
757 .sr(1)
758 .m(1)
759 .n(n)
760 .k(k)
761 .cn_stride(11)
762 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
763 }
764 }
765 }
766
767 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_subtile) {
768 TEST_REQUIRES_ARM_NEON;
769 for (uint32_t n = 16; n <= 24; n += 8) {
770 for (size_t k = 1; k <= 40; k += 9) {
771 for (uint32_t m = 1; m <= 1; m++) {
772 GemmMicrokernelTester()
773 .mr(1)
774 .nr(8)
775 .kr(2)
776 .sr(1)
777 .m(m)
778 .n(n)
779 .k(k)
780 .iterations(1)
781 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
782 }
783 }
784 }
785 }
786
787 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, small_kernel) {
788 TEST_REQUIRES_ARM_NEON;
789 for (size_t k = 1; k <= 40; k += 9) {
790 GemmMicrokernelTester()
791 .mr(1)
792 .nr(8)
793 .kr(2)
794 .sr(1)
795 .m(1)
796 .n(8)
797 .k(k)
798 .ks(3)
799 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
800 }
801 }
802
803 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, small_kernel_subtile) {
804 TEST_REQUIRES_ARM_NEON;
805 for (size_t k = 1; k <= 40; k += 9) {
806 for (uint32_t n = 1; n <= 8; n++) {
807 for (uint32_t m = 1; m <= 1; m++) {
808 GemmMicrokernelTester()
809 .mr(1)
810 .nr(8)
811 .kr(2)
812 .sr(1)
813 .m(m)
814 .n(n)
815 .k(k)
816 .ks(3)
817 .iterations(1)
818 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
819 }
820 }
821 }
822 }
823
824 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_gt_8_small_kernel) {
825 TEST_REQUIRES_ARM_NEON;
826 for (uint32_t n = 9; n < 16; n++) {
827 for (size_t k = 1; k <= 40; k += 9) {
828 GemmMicrokernelTester()
829 .mr(1)
830 .nr(8)
831 .kr(2)
832 .sr(1)
833 .m(1)
834 .n(n)
835 .k(k)
836 .ks(3)
837 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
838 }
839 }
840 }
841
842 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, n_div_8_small_kernel) {
843 TEST_REQUIRES_ARM_NEON;
844 for (uint32_t n = 16; n <= 24; n += 8) {
845 for (size_t k = 1; k <= 40; k += 9) {
846 GemmMicrokernelTester()
847 .mr(1)
848 .nr(8)
849 .kr(2)
850 .sr(1)
851 .m(1)
852 .n(n)
853 .k(k)
854 .ks(3)
855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
856 }
857 }
858 }
859
860 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, strided_cm_subtile) {
861 TEST_REQUIRES_ARM_NEON;
862 for (size_t k = 1; k <= 40; k += 9) {
863 for (uint32_t n = 1; n <= 8; n++) {
864 for (uint32_t m = 1; m <= 1; m++) {
865 GemmMicrokernelTester()
866 .mr(1)
867 .nr(8)
868 .kr(2)
869 .sr(1)
870 .m(m)
871 .n(n)
872 .k(k)
873 .cm_stride(11)
874 .iterations(1)
875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
876 }
877 }
878 }
879 }
880
881 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, a_offset) {
882 TEST_REQUIRES_ARM_NEON;
883 for (size_t k = 1; k <= 40; k += 9) {
884 GemmMicrokernelTester()
885 .mr(1)
886 .nr(8)
887 .kr(2)
888 .sr(1)
889 .m(1)
890 .n(8)
891 .k(k)
892 .ks(3)
893 .a_offset(43)
894 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
895 }
896 }
897
898 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, zero) {
899 TEST_REQUIRES_ARM_NEON;
900 for (size_t k = 1; k <= 40; k += 9) {
901 for (uint32_t mz = 0; mz < 1; mz++) {
902 GemmMicrokernelTester()
903 .mr(1)
904 .nr(8)
905 .kr(2)
906 .sr(1)
907 .m(1)
908 .n(8)
909 .k(k)
910 .ks(3)
911 .a_offset(43)
912 .zero_index(mz)
913 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
914 }
915 }
916 }
917
918 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, qmin) {
919 TEST_REQUIRES_ARM_NEON;
920 GemmMicrokernelTester()
921 .mr(1)
922 .nr(8)
923 .kr(2)
924 .sr(1)
925 .m(1)
926 .n(8)
927 .k(8)
928 .qmin(128)
929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
930 }
931
932 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, qmax) {
933 TEST_REQUIRES_ARM_NEON;
934 GemmMicrokernelTester()
935 .mr(1)
936 .nr(8)
937 .kr(2)
938 .sr(1)
939 .m(1)
940 .n(8)
941 .k(8)
942 .qmax(128)
943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
944 }
945
946 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MULL_LD1R, strided_cm) {
947 TEST_REQUIRES_ARM_NEON;
948 GemmMicrokernelTester()
949 .mr(1)
950 .nr(8)
951 .kr(2)
952 .sr(1)
953 .m(1)
954 .n(8)
955 .k(8)
956 .cm_stride(11)
957 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
958 }
959#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
960
961
962#if XNN_ARCH_ARM || XNN_ARCH_ARM64
963 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_eq_16) {
964 TEST_REQUIRES_ARM_NEON;
965 GemmMicrokernelTester()
966 .mr(1)
967 .nr(8)
968 .kr(2)
969 .sr(1)
970 .m(1)
971 .n(8)
972 .k(16)
973 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
974 }
975
976 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, strided_cn) {
977 TEST_REQUIRES_ARM_NEON;
978 GemmMicrokernelTester()
979 .mr(1)
980 .nr(8)
981 .kr(2)
982 .sr(1)
983 .m(1)
984 .n(8)
985 .k(16)
986 .cn_stride(11)
987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
988 }
989
990 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_eq_16_subtile) {
991 TEST_REQUIRES_ARM_NEON;
992 for (uint32_t n = 1; n <= 8; n++) {
993 for (uint32_t m = 1; m <= 1; m++) {
994 GemmMicrokernelTester()
995 .mr(1)
996 .nr(8)
997 .kr(2)
998 .sr(1)
999 .m(m)
1000 .n(n)
1001 .k(16)
1002 .iterations(1)
1003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1004 }
1005 }
1006 }
1007
1008 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
1009 TEST_REQUIRES_ARM_NEON;
1010 for (uint32_t m = 1; m <= 1; m++) {
1011 GemmMicrokernelTester()
1012 .mr(1)
1013 .nr(8)
1014 .kr(2)
1015 .sr(1)
1016 .m(m)
1017 .n(8)
1018 .k(16)
1019 .iterations(1)
1020 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1021 }
1022 }
1023
1024 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
1025 TEST_REQUIRES_ARM_NEON;
1026 for (uint32_t n = 1; n <= 8; n++) {
1027 GemmMicrokernelTester()
1028 .mr(1)
1029 .nr(8)
1030 .kr(2)
1031 .sr(1)
1032 .m(1)
1033 .n(n)
1034 .k(16)
1035 .iterations(1)
1036 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1037 }
1038 }
1039
1040 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_lt_16) {
1041 TEST_REQUIRES_ARM_NEON;
1042 for (size_t k = 1; k < 16; k++) {
1043 GemmMicrokernelTester()
1044 .mr(1)
1045 .nr(8)
1046 .kr(2)
1047 .sr(1)
1048 .m(1)
1049 .n(8)
1050 .k(k)
1051 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1052 }
1053 }
1054
1055 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_lt_16_subtile) {
1056 TEST_REQUIRES_ARM_NEON;
1057 for (size_t k = 1; k < 16; k++) {
1058 for (uint32_t n = 1; n <= 8; n++) {
1059 for (uint32_t m = 1; m <= 1; m++) {
1060 GemmMicrokernelTester()
1061 .mr(1)
1062 .nr(8)
1063 .kr(2)
1064 .sr(1)
1065 .m(m)
1066 .n(n)
1067 .k(k)
1068 .iterations(1)
1069 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1070 }
1071 }
1072 }
1073 }
1074
1075 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_gt_16) {
1076 TEST_REQUIRES_ARM_NEON;
1077 for (size_t k = 17; k < 32; k++) {
1078 GemmMicrokernelTester()
1079 .mr(1)
1080 .nr(8)
1081 .kr(2)
1082 .sr(1)
1083 .m(1)
1084 .n(8)
1085 .k(k)
1086 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1087 }
1088 }
1089
1090 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_gt_16_subtile) {
1091 TEST_REQUIRES_ARM_NEON;
1092 for (size_t k = 17; k < 32; k++) {
1093 for (uint32_t n = 1; n <= 8; n++) {
1094 for (uint32_t m = 1; m <= 1; m++) {
1095 GemmMicrokernelTester()
1096 .mr(1)
1097 .nr(8)
1098 .kr(2)
1099 .sr(1)
1100 .m(m)
1101 .n(n)
1102 .k(k)
1103 .iterations(1)
1104 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1105 }
1106 }
1107 }
1108 }
1109
1110 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_div_16) {
1111 TEST_REQUIRES_ARM_NEON;
1112 for (size_t k = 32; k <= 160; k += 16) {
1113 GemmMicrokernelTester()
1114 .mr(1)
1115 .nr(8)
1116 .kr(2)
1117 .sr(1)
1118 .m(1)
1119 .n(8)
1120 .k(k)
1121 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1122 }
1123 }
1124
1125 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, k_div_16_subtile) {
1126 TEST_REQUIRES_ARM_NEON;
1127 for (size_t k = 32; k <= 160; k += 16) {
1128 for (uint32_t n = 1; n <= 8; n++) {
1129 for (uint32_t m = 1; m <= 1; m++) {
1130 GemmMicrokernelTester()
1131 .mr(1)
1132 .nr(8)
1133 .kr(2)
1134 .sr(1)
1135 .m(m)
1136 .n(n)
1137 .k(k)
1138 .iterations(1)
1139 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1140 }
1141 }
1142 }
1143 }
1144
1145 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_gt_8) {
1146 TEST_REQUIRES_ARM_NEON;
1147 for (uint32_t n = 9; n < 16; n++) {
1148 for (size_t k = 1; k <= 80; k += 17) {
1149 GemmMicrokernelTester()
1150 .mr(1)
1151 .nr(8)
1152 .kr(2)
1153 .sr(1)
1154 .m(1)
1155 .n(n)
1156 .k(k)
1157 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1158 }
1159 }
1160 }
1161
1162 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_gt_8_strided_cn) {
1163 TEST_REQUIRES_ARM_NEON;
1164 for (uint32_t n = 9; n < 16; n++) {
1165 for (size_t k = 1; k <= 80; k += 17) {
1166 GemmMicrokernelTester()
1167 .mr(1)
1168 .nr(8)
1169 .kr(2)
1170 .sr(1)
1171 .m(1)
1172 .n(n)
1173 .k(k)
1174 .cn_stride(11)
1175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1176 }
1177 }
1178 }
1179
1180 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_gt_8_subtile) {
1181 TEST_REQUIRES_ARM_NEON;
1182 for (uint32_t n = 9; n < 16; n++) {
1183 for (size_t k = 1; k <= 80; k += 17) {
1184 for (uint32_t m = 1; m <= 1; m++) {
1185 GemmMicrokernelTester()
1186 .mr(1)
1187 .nr(8)
1188 .kr(2)
1189 .sr(1)
1190 .m(m)
1191 .n(n)
1192 .k(k)
1193 .iterations(1)
1194 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1195 }
1196 }
1197 }
1198 }
1199
1200 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_div_8) {
1201 TEST_REQUIRES_ARM_NEON;
1202 for (uint32_t n = 16; n <= 24; n += 8) {
1203 for (size_t k = 1; k <= 80; k += 17) {
1204 GemmMicrokernelTester()
1205 .mr(1)
1206 .nr(8)
1207 .kr(2)
1208 .sr(1)
1209 .m(1)
1210 .n(n)
1211 .k(k)
1212 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1213 }
1214 }
1215 }
1216
1217 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_div_8_strided_cn) {
1218 TEST_REQUIRES_ARM_NEON;
1219 for (uint32_t n = 16; n <= 24; n += 8) {
1220 for (size_t k = 1; k <= 80; k += 17) {
1221 GemmMicrokernelTester()
1222 .mr(1)
1223 .nr(8)
1224 .kr(2)
1225 .sr(1)
1226 .m(1)
1227 .n(n)
1228 .k(k)
1229 .cn_stride(11)
1230 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1231 }
1232 }
1233 }
1234
1235 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_div_8_subtile) {
1236 TEST_REQUIRES_ARM_NEON;
1237 for (uint32_t n = 16; n <= 24; n += 8) {
1238 for (size_t k = 1; k <= 80; k += 17) {
1239 for (uint32_t m = 1; m <= 1; m++) {
1240 GemmMicrokernelTester()
1241 .mr(1)
1242 .nr(8)
1243 .kr(2)
1244 .sr(1)
1245 .m(m)
1246 .n(n)
1247 .k(k)
1248 .iterations(1)
1249 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1250 }
1251 }
1252 }
1253 }
1254
1255 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, small_kernel) {
1256 TEST_REQUIRES_ARM_NEON;
1257 for (size_t k = 1; k <= 80; k += 17) {
1258 GemmMicrokernelTester()
1259 .mr(1)
1260 .nr(8)
1261 .kr(2)
1262 .sr(1)
1263 .m(1)
1264 .n(8)
1265 .k(k)
1266 .ks(3)
1267 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1268 }
1269 }
1270
1271 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, small_kernel_subtile) {
1272 TEST_REQUIRES_ARM_NEON;
1273 for (size_t k = 1; k <= 80; k += 17) {
1274 for (uint32_t n = 1; n <= 8; n++) {
1275 for (uint32_t m = 1; m <= 1; m++) {
1276 GemmMicrokernelTester()
1277 .mr(1)
1278 .nr(8)
1279 .kr(2)
1280 .sr(1)
1281 .m(m)
1282 .n(n)
1283 .k(k)
1284 .ks(3)
1285 .iterations(1)
1286 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1287 }
1288 }
1289 }
1290 }
1291
1292 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_gt_8_small_kernel) {
1293 TEST_REQUIRES_ARM_NEON;
1294 for (uint32_t n = 9; n < 16; n++) {
1295 for (size_t k = 1; k <= 80; k += 17) {
1296 GemmMicrokernelTester()
1297 .mr(1)
1298 .nr(8)
1299 .kr(2)
1300 .sr(1)
1301 .m(1)
1302 .n(n)
1303 .k(k)
1304 .ks(3)
1305 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1306 }
1307 }
1308 }
1309
1310 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, n_div_8_small_kernel) {
1311 TEST_REQUIRES_ARM_NEON;
1312 for (uint32_t n = 16; n <= 24; n += 8) {
1313 for (size_t k = 1; k <= 80; k += 17) {
1314 GemmMicrokernelTester()
1315 .mr(1)
1316 .nr(8)
1317 .kr(2)
1318 .sr(1)
1319 .m(1)
1320 .n(n)
1321 .k(k)
1322 .ks(3)
1323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1324 }
1325 }
1326 }
1327
1328 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, strided_cm_subtile) {
1329 TEST_REQUIRES_ARM_NEON;
1330 for (size_t k = 1; k <= 80; k += 17) {
1331 for (uint32_t n = 1; n <= 8; n++) {
1332 for (uint32_t m = 1; m <= 1; m++) {
1333 GemmMicrokernelTester()
1334 .mr(1)
1335 .nr(8)
1336 .kr(2)
1337 .sr(1)
1338 .m(m)
1339 .n(n)
1340 .k(k)
1341 .cm_stride(11)
1342 .iterations(1)
1343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1344 }
1345 }
1346 }
1347 }
1348
1349 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, a_offset) {
1350 TEST_REQUIRES_ARM_NEON;
1351 for (size_t k = 1; k <= 80; k += 17) {
1352 GemmMicrokernelTester()
1353 .mr(1)
1354 .nr(8)
1355 .kr(2)
1356 .sr(1)
1357 .m(1)
1358 .n(8)
1359 .k(k)
1360 .ks(3)
1361 .a_offset(83)
1362 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1363 }
1364 }
1365
1366 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, zero) {
1367 TEST_REQUIRES_ARM_NEON;
1368 for (size_t k = 1; k <= 80; k += 17) {
1369 for (uint32_t mz = 0; mz < 1; mz++) {
1370 GemmMicrokernelTester()
1371 .mr(1)
1372 .nr(8)
1373 .kr(2)
1374 .sr(1)
1375 .m(1)
1376 .n(8)
1377 .k(k)
1378 .ks(3)
1379 .a_offset(83)
1380 .zero_index(mz)
1381 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1382 }
1383 }
1384 }
1385
1386 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, qmin) {
1387 TEST_REQUIRES_ARM_NEON;
1388 GemmMicrokernelTester()
1389 .mr(1)
1390 .nr(8)
1391 .kr(2)
1392 .sr(1)
1393 .m(1)
1394 .n(8)
1395 .k(16)
1396 .qmin(128)
1397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1398 }
1399
1400 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, qmax) {
1401 TEST_REQUIRES_ARM_NEON;
1402 GemmMicrokernelTester()
1403 .mr(1)
1404 .nr(8)
1405 .kr(2)
1406 .sr(1)
1407 .m(1)
1408 .n(8)
1409 .k(16)
1410 .qmax(128)
1411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1412 }
1413
1414 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_LD1R, strided_cm) {
1415 TEST_REQUIRES_ARM_NEON;
1416 GemmMicrokernelTester()
1417 .mr(1)
1418 .nr(8)
1419 .kr(2)
1420 .sr(1)
1421 .m(1)
1422 .n(8)
1423 .k(16)
1424 .cm_stride(11)
1425 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1426 }
1427#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
1428
1429
1430#if XNN_ARCH_ARM || XNN_ARCH_ARM64
1431 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8) {
1432 TEST_REQUIRES_ARM_NEON;
1433 GemmMicrokernelTester()
1434 .mr(2)
1435 .nr(8)
1436 .kr(2)
1437 .sr(1)
1438 .m(2)
1439 .n(8)
1440 .k(8)
1441 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1442 }
1443
1444 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, strided_cn) {
1445 TEST_REQUIRES_ARM_NEON;
1446 GemmMicrokernelTester()
1447 .mr(2)
1448 .nr(8)
1449 .kr(2)
1450 .sr(1)
1451 .m(2)
1452 .n(8)
1453 .k(8)
1454 .cn_stride(11)
1455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1456 }
1457
1458 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8_subtile) {
1459 TEST_REQUIRES_ARM_NEON;
1460 for (uint32_t n = 1; n <= 8; n++) {
1461 for (uint32_t m = 1; m <= 2; m++) {
1462 GemmMicrokernelTester()
1463 .mr(2)
1464 .nr(8)
1465 .kr(2)
1466 .sr(1)
1467 .m(m)
1468 .n(n)
1469 .k(8)
1470 .iterations(1)
1471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1472 }
1473 }
1474 }
1475
1476 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8_subtile_m) {
1477 TEST_REQUIRES_ARM_NEON;
1478 for (uint32_t m = 1; m <= 2; m++) {
1479 GemmMicrokernelTester()
1480 .mr(2)
1481 .nr(8)
1482 .kr(2)
1483 .sr(1)
1484 .m(m)
1485 .n(8)
1486 .k(8)
1487 .iterations(1)
1488 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1489 }
1490 }
1491
1492 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_eq_8_subtile_n) {
1493 TEST_REQUIRES_ARM_NEON;
1494 for (uint32_t n = 1; n <= 8; n++) {
1495 GemmMicrokernelTester()
1496 .mr(2)
1497 .nr(8)
1498 .kr(2)
1499 .sr(1)
1500 .m(2)
1501 .n(n)
1502 .k(8)
1503 .iterations(1)
1504 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1505 }
1506 }
1507
1508 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_lt_8) {
1509 TEST_REQUIRES_ARM_NEON;
1510 for (size_t k = 1; k < 8; k++) {
1511 GemmMicrokernelTester()
1512 .mr(2)
1513 .nr(8)
1514 .kr(2)
1515 .sr(1)
1516 .m(2)
1517 .n(8)
1518 .k(k)
1519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1520 }
1521 }
1522
1523 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_lt_8_subtile) {
1524 TEST_REQUIRES_ARM_NEON;
1525 for (size_t k = 1; k < 8; k++) {
1526 for (uint32_t n = 1; n <= 8; n++) {
1527 for (uint32_t m = 1; m <= 2; m++) {
1528 GemmMicrokernelTester()
1529 .mr(2)
1530 .nr(8)
1531 .kr(2)
1532 .sr(1)
1533 .m(m)
1534 .n(n)
1535 .k(k)
1536 .iterations(1)
1537 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1538 }
1539 }
1540 }
1541 }
1542
1543 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_gt_8) {
1544 TEST_REQUIRES_ARM_NEON;
1545 for (size_t k = 9; k < 16; k++) {
1546 GemmMicrokernelTester()
1547 .mr(2)
1548 .nr(8)
1549 .kr(2)
1550 .sr(1)
1551 .m(2)
1552 .n(8)
1553 .k(k)
1554 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1555 }
1556 }
1557
1558 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_gt_8_subtile) {
1559 TEST_REQUIRES_ARM_NEON;
1560 for (size_t k = 9; k < 16; k++) {
1561 for (uint32_t n = 1; n <= 8; n++) {
1562 for (uint32_t m = 1; m <= 2; m++) {
1563 GemmMicrokernelTester()
1564 .mr(2)
1565 .nr(8)
1566 .kr(2)
1567 .sr(1)
1568 .m(m)
1569 .n(n)
1570 .k(k)
1571 .iterations(1)
1572 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1573 }
1574 }
1575 }
1576 }
1577
1578 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_div_8) {
1579 TEST_REQUIRES_ARM_NEON;
1580 for (size_t k = 16; k <= 80; k += 8) {
1581 GemmMicrokernelTester()
1582 .mr(2)
1583 .nr(8)
1584 .kr(2)
1585 .sr(1)
1586 .m(2)
1587 .n(8)
1588 .k(k)
1589 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1590 }
1591 }
1592
1593 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, k_div_8_subtile) {
1594 TEST_REQUIRES_ARM_NEON;
1595 for (size_t k = 16; k <= 80; k += 8) {
1596 for (uint32_t n = 1; n <= 8; n++) {
1597 for (uint32_t m = 1; m <= 2; m++) {
1598 GemmMicrokernelTester()
1599 .mr(2)
1600 .nr(8)
1601 .kr(2)
1602 .sr(1)
1603 .m(m)
1604 .n(n)
1605 .k(k)
1606 .iterations(1)
1607 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1608 }
1609 }
1610 }
1611 }
1612
1613 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8) {
1614 TEST_REQUIRES_ARM_NEON;
1615 for (uint32_t n = 9; n < 16; n++) {
1616 for (size_t k = 1; k <= 40; k += 9) {
1617 GemmMicrokernelTester()
1618 .mr(2)
1619 .nr(8)
1620 .kr(2)
1621 .sr(1)
1622 .m(2)
1623 .n(n)
1624 .k(k)
1625 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1626 }
1627 }
1628 }
1629
1630 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8_strided_cn) {
1631 TEST_REQUIRES_ARM_NEON;
1632 for (uint32_t n = 9; n < 16; n++) {
1633 for (size_t k = 1; k <= 40; k += 9) {
1634 GemmMicrokernelTester()
1635 .mr(2)
1636 .nr(8)
1637 .kr(2)
1638 .sr(1)
1639 .m(2)
1640 .n(n)
1641 .k(k)
1642 .cn_stride(11)
1643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1644 }
1645 }
1646 }
1647
1648 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8_subtile) {
1649 TEST_REQUIRES_ARM_NEON;
1650 for (uint32_t n = 9; n < 16; n++) {
1651 for (size_t k = 1; k <= 40; k += 9) {
1652 for (uint32_t m = 1; m <= 2; m++) {
1653 GemmMicrokernelTester()
1654 .mr(2)
1655 .nr(8)
1656 .kr(2)
1657 .sr(1)
1658 .m(m)
1659 .n(n)
1660 .k(k)
1661 .iterations(1)
1662 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1663 }
1664 }
1665 }
1666 }
1667
1668 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8) {
1669 TEST_REQUIRES_ARM_NEON;
1670 for (uint32_t n = 16; n <= 24; n += 8) {
1671 for (size_t k = 1; k <= 40; k += 9) {
1672 GemmMicrokernelTester()
1673 .mr(2)
1674 .nr(8)
1675 .kr(2)
1676 .sr(1)
1677 .m(2)
1678 .n(n)
1679 .k(k)
1680 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1681 }
1682 }
1683 }
1684
1685 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8_strided_cn) {
1686 TEST_REQUIRES_ARM_NEON;
1687 for (uint32_t n = 16; n <= 24; n += 8) {
1688 for (size_t k = 1; k <= 40; k += 9) {
1689 GemmMicrokernelTester()
1690 .mr(2)
1691 .nr(8)
1692 .kr(2)
1693 .sr(1)
1694 .m(2)
1695 .n(n)
1696 .k(k)
1697 .cn_stride(11)
1698 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1699 }
1700 }
1701 }
1702
1703 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8_subtile) {
1704 TEST_REQUIRES_ARM_NEON;
1705 for (uint32_t n = 16; n <= 24; n += 8) {
1706 for (size_t k = 1; k <= 40; k += 9) {
1707 for (uint32_t m = 1; m <= 2; m++) {
1708 GemmMicrokernelTester()
1709 .mr(2)
1710 .nr(8)
1711 .kr(2)
1712 .sr(1)
1713 .m(m)
1714 .n(n)
1715 .k(k)
1716 .iterations(1)
1717 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1718 }
1719 }
1720 }
1721 }
1722
1723 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, small_kernel) {
1724 TEST_REQUIRES_ARM_NEON;
1725 for (size_t k = 1; k <= 40; k += 9) {
1726 GemmMicrokernelTester()
1727 .mr(2)
1728 .nr(8)
1729 .kr(2)
1730 .sr(1)
1731 .m(2)
1732 .n(8)
1733 .k(k)
1734 .ks(3)
1735 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1736 }
1737 }
1738
1739 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, small_kernel_subtile) {
1740 TEST_REQUIRES_ARM_NEON;
1741 for (size_t k = 1; k <= 40; k += 9) {
1742 for (uint32_t n = 1; n <= 8; n++) {
1743 for (uint32_t m = 1; m <= 2; m++) {
1744 GemmMicrokernelTester()
1745 .mr(2)
1746 .nr(8)
1747 .kr(2)
1748 .sr(1)
1749 .m(m)
1750 .n(n)
1751 .k(k)
1752 .ks(3)
1753 .iterations(1)
1754 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1755 }
1756 }
1757 }
1758 }
1759
1760 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_gt_8_small_kernel) {
1761 TEST_REQUIRES_ARM_NEON;
1762 for (uint32_t n = 9; n < 16; n++) {
1763 for (size_t k = 1; k <= 40; k += 9) {
1764 GemmMicrokernelTester()
1765 .mr(2)
1766 .nr(8)
1767 .kr(2)
1768 .sr(1)
1769 .m(2)
1770 .n(n)
1771 .k(k)
1772 .ks(3)
1773 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1774 }
1775 }
1776 }
1777
1778 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, n_div_8_small_kernel) {
1779 TEST_REQUIRES_ARM_NEON;
1780 for (uint32_t n = 16; n <= 24; n += 8) {
1781 for (size_t k = 1; k <= 40; k += 9) {
1782 GemmMicrokernelTester()
1783 .mr(2)
1784 .nr(8)
1785 .kr(2)
1786 .sr(1)
1787 .m(2)
1788 .n(n)
1789 .k(k)
1790 .ks(3)
1791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1792 }
1793 }
1794 }
1795
1796 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, strided_cm_subtile) {
1797 TEST_REQUIRES_ARM_NEON;
1798 for (size_t k = 1; k <= 40; k += 9) {
1799 for (uint32_t n = 1; n <= 8; n++) {
1800 for (uint32_t m = 1; m <= 2; m++) {
1801 GemmMicrokernelTester()
1802 .mr(2)
1803 .nr(8)
1804 .kr(2)
1805 .sr(1)
1806 .m(m)
1807 .n(n)
1808 .k(k)
1809 .cm_stride(11)
1810 .iterations(1)
1811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1812 }
1813 }
1814 }
1815 }
1816
1817 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, a_offset) {
1818 TEST_REQUIRES_ARM_NEON;
1819 for (size_t k = 1; k <= 40; k += 9) {
1820 GemmMicrokernelTester()
1821 .mr(2)
1822 .nr(8)
1823 .kr(2)
1824 .sr(1)
1825 .m(2)
1826 .n(8)
1827 .k(k)
1828 .ks(3)
1829 .a_offset(83)
1830 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1831 }
1832 }
1833
1834 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, zero) {
1835 TEST_REQUIRES_ARM_NEON;
1836 for (size_t k = 1; k <= 40; k += 9) {
1837 for (uint32_t mz = 0; mz < 2; mz++) {
1838 GemmMicrokernelTester()
1839 .mr(2)
1840 .nr(8)
1841 .kr(2)
1842 .sr(1)
1843 .m(2)
1844 .n(8)
1845 .k(k)
1846 .ks(3)
1847 .a_offset(83)
1848 .zero_index(mz)
1849 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1850 }
1851 }
1852 }
1853
1854 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, qmin) {
1855 TEST_REQUIRES_ARM_NEON;
1856 GemmMicrokernelTester()
1857 .mr(2)
1858 .nr(8)
1859 .kr(2)
1860 .sr(1)
1861 .m(2)
1862 .n(8)
1863 .k(8)
1864 .qmin(128)
1865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1866 }
1867
1868 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, qmax) {
1869 TEST_REQUIRES_ARM_NEON;
1870 GemmMicrokernelTester()
1871 .mr(2)
1872 .nr(8)
1873 .kr(2)
1874 .sr(1)
1875 .m(2)
1876 .n(8)
1877 .k(8)
1878 .qmax(128)
1879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1880 }
1881
1882 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MULL_LD2R, strided_cm) {
1883 TEST_REQUIRES_ARM_NEON;
1884 GemmMicrokernelTester()
1885 .mr(2)
1886 .nr(8)
1887 .kr(2)
1888 .sr(1)
1889 .m(2)
1890 .n(8)
1891 .k(8)
1892 .cm_stride(11)
1893 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1894 }
1895#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
1896
1897
1898#if XNN_ARCH_ARM || XNN_ARCH_ARM64
1899 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_eq_16) {
1900 TEST_REQUIRES_ARM_NEON;
1901 GemmMicrokernelTester()
1902 .mr(2)
1903 .nr(8)
1904 .kr(2)
1905 .sr(1)
1906 .m(2)
1907 .n(8)
1908 .k(16)
1909 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1910 }
1911
1912 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, strided_cn) {
1913 TEST_REQUIRES_ARM_NEON;
1914 GemmMicrokernelTester()
1915 .mr(2)
1916 .nr(8)
1917 .kr(2)
1918 .sr(1)
1919 .m(2)
1920 .n(8)
1921 .k(16)
1922 .cn_stride(11)
1923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1924 }
1925
1926 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_eq_16_subtile) {
1927 TEST_REQUIRES_ARM_NEON;
1928 for (uint32_t n = 1; n <= 8; n++) {
1929 for (uint32_t m = 1; m <= 2; m++) {
1930 GemmMicrokernelTester()
1931 .mr(2)
1932 .nr(8)
1933 .kr(2)
1934 .sr(1)
1935 .m(m)
1936 .n(n)
1937 .k(16)
1938 .iterations(1)
1939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1940 }
1941 }
1942 }
1943
1944 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_eq_16_subtile_m) {
1945 TEST_REQUIRES_ARM_NEON;
1946 for (uint32_t m = 1; m <= 2; m++) {
1947 GemmMicrokernelTester()
1948 .mr(2)
1949 .nr(8)
1950 .kr(2)
1951 .sr(1)
1952 .m(m)
1953 .n(8)
1954 .k(16)
1955 .iterations(1)
1956 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1957 }
1958 }
1959
1960 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_eq_16_subtile_n) {
1961 TEST_REQUIRES_ARM_NEON;
1962 for (uint32_t n = 1; n <= 8; n++) {
1963 GemmMicrokernelTester()
1964 .mr(2)
1965 .nr(8)
1966 .kr(2)
1967 .sr(1)
1968 .m(2)
1969 .n(n)
1970 .k(16)
1971 .iterations(1)
1972 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1973 }
1974 }
1975
1976 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_lt_16) {
1977 TEST_REQUIRES_ARM_NEON;
1978 for (size_t k = 1; k < 16; k++) {
1979 GemmMicrokernelTester()
1980 .mr(2)
1981 .nr(8)
1982 .kr(2)
1983 .sr(1)
1984 .m(2)
1985 .n(8)
1986 .k(k)
1987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
1988 }
1989 }
1990
1991 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_lt_16_subtile) {
1992 TEST_REQUIRES_ARM_NEON;
1993 for (size_t k = 1; k < 16; k++) {
1994 for (uint32_t n = 1; n <= 8; n++) {
1995 for (uint32_t m = 1; m <= 2; m++) {
1996 GemmMicrokernelTester()
1997 .mr(2)
1998 .nr(8)
1999 .kr(2)
2000 .sr(1)
2001 .m(m)
2002 .n(n)
2003 .k(k)
2004 .iterations(1)
2005 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2006 }
2007 }
2008 }
2009 }
2010
2011 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_gt_16) {
2012 TEST_REQUIRES_ARM_NEON;
2013 for (size_t k = 17; k < 32; k++) {
2014 GemmMicrokernelTester()
2015 .mr(2)
2016 .nr(8)
2017 .kr(2)
2018 .sr(1)
2019 .m(2)
2020 .n(8)
2021 .k(k)
2022 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2023 }
2024 }
2025
2026 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_gt_16_subtile) {
2027 TEST_REQUIRES_ARM_NEON;
2028 for (size_t k = 17; k < 32; k++) {
2029 for (uint32_t n = 1; n <= 8; n++) {
2030 for (uint32_t m = 1; m <= 2; m++) {
2031 GemmMicrokernelTester()
2032 .mr(2)
2033 .nr(8)
2034 .kr(2)
2035 .sr(1)
2036 .m(m)
2037 .n(n)
2038 .k(k)
2039 .iterations(1)
2040 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2041 }
2042 }
2043 }
2044 }
2045
2046 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_div_16) {
2047 TEST_REQUIRES_ARM_NEON;
2048 for (size_t k = 32; k <= 160; k += 16) {
2049 GemmMicrokernelTester()
2050 .mr(2)
2051 .nr(8)
2052 .kr(2)
2053 .sr(1)
2054 .m(2)
2055 .n(8)
2056 .k(k)
2057 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2058 }
2059 }
2060
2061 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, k_div_16_subtile) {
2062 TEST_REQUIRES_ARM_NEON;
2063 for (size_t k = 32; k <= 160; k += 16) {
2064 for (uint32_t n = 1; n <= 8; n++) {
2065 for (uint32_t m = 1; m <= 2; m++) {
2066 GemmMicrokernelTester()
2067 .mr(2)
2068 .nr(8)
2069 .kr(2)
2070 .sr(1)
2071 .m(m)
2072 .n(n)
2073 .k(k)
2074 .iterations(1)
2075 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2076 }
2077 }
2078 }
2079 }
2080
2081 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_gt_8) {
2082 TEST_REQUIRES_ARM_NEON;
2083 for (uint32_t n = 9; n < 16; n++) {
2084 for (size_t k = 1; k <= 80; k += 17) {
2085 GemmMicrokernelTester()
2086 .mr(2)
2087 .nr(8)
2088 .kr(2)
2089 .sr(1)
2090 .m(2)
2091 .n(n)
2092 .k(k)
2093 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2094 }
2095 }
2096 }
2097
2098 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_gt_8_strided_cn) {
2099 TEST_REQUIRES_ARM_NEON;
2100 for (uint32_t n = 9; n < 16; n++) {
2101 for (size_t k = 1; k <= 80; k += 17) {
2102 GemmMicrokernelTester()
2103 .mr(2)
2104 .nr(8)
2105 .kr(2)
2106 .sr(1)
2107 .m(2)
2108 .n(n)
2109 .k(k)
2110 .cn_stride(11)
2111 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2112 }
2113 }
2114 }
2115
2116 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_gt_8_subtile) {
2117 TEST_REQUIRES_ARM_NEON;
2118 for (uint32_t n = 9; n < 16; n++) {
2119 for (size_t k = 1; k <= 80; k += 17) {
2120 for (uint32_t m = 1; m <= 2; m++) {
2121 GemmMicrokernelTester()
2122 .mr(2)
2123 .nr(8)
2124 .kr(2)
2125 .sr(1)
2126 .m(m)
2127 .n(n)
2128 .k(k)
2129 .iterations(1)
2130 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2131 }
2132 }
2133 }
2134 }
2135
2136 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_div_8) {
2137 TEST_REQUIRES_ARM_NEON;
2138 for (uint32_t n = 16; n <= 24; n += 8) {
2139 for (size_t k = 1; k <= 80; k += 17) {
2140 GemmMicrokernelTester()
2141 .mr(2)
2142 .nr(8)
2143 .kr(2)
2144 .sr(1)
2145 .m(2)
2146 .n(n)
2147 .k(k)
2148 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2149 }
2150 }
2151 }
2152
2153 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_div_8_strided_cn) {
2154 TEST_REQUIRES_ARM_NEON;
2155 for (uint32_t n = 16; n <= 24; n += 8) {
2156 for (size_t k = 1; k <= 80; k += 17) {
2157 GemmMicrokernelTester()
2158 .mr(2)
2159 .nr(8)
2160 .kr(2)
2161 .sr(1)
2162 .m(2)
2163 .n(n)
2164 .k(k)
2165 .cn_stride(11)
2166 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2167 }
2168 }
2169 }
2170
2171 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_div_8_subtile) {
2172 TEST_REQUIRES_ARM_NEON;
2173 for (uint32_t n = 16; n <= 24; n += 8) {
2174 for (size_t k = 1; k <= 80; k += 17) {
2175 for (uint32_t m = 1; m <= 2; m++) {
2176 GemmMicrokernelTester()
2177 .mr(2)
2178 .nr(8)
2179 .kr(2)
2180 .sr(1)
2181 .m(m)
2182 .n(n)
2183 .k(k)
2184 .iterations(1)
2185 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2186 }
2187 }
2188 }
2189 }
2190
2191 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, small_kernel) {
2192 TEST_REQUIRES_ARM_NEON;
2193 for (size_t k = 1; k <= 80; k += 17) {
2194 GemmMicrokernelTester()
2195 .mr(2)
2196 .nr(8)
2197 .kr(2)
2198 .sr(1)
2199 .m(2)
2200 .n(8)
2201 .k(k)
2202 .ks(3)
2203 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2204 }
2205 }
2206
2207 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, small_kernel_subtile) {
2208 TEST_REQUIRES_ARM_NEON;
2209 for (size_t k = 1; k <= 80; k += 17) {
2210 for (uint32_t n = 1; n <= 8; n++) {
2211 for (uint32_t m = 1; m <= 2; m++) {
2212 GemmMicrokernelTester()
2213 .mr(2)
2214 .nr(8)
2215 .kr(2)
2216 .sr(1)
2217 .m(m)
2218 .n(n)
2219 .k(k)
2220 .ks(3)
2221 .iterations(1)
2222 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2223 }
2224 }
2225 }
2226 }
2227
2228 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_gt_8_small_kernel) {
2229 TEST_REQUIRES_ARM_NEON;
2230 for (uint32_t n = 9; n < 16; n++) {
2231 for (size_t k = 1; k <= 80; k += 17) {
2232 GemmMicrokernelTester()
2233 .mr(2)
2234 .nr(8)
2235 .kr(2)
2236 .sr(1)
2237 .m(2)
2238 .n(n)
2239 .k(k)
2240 .ks(3)
2241 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2242 }
2243 }
2244 }
2245
2246 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, n_div_8_small_kernel) {
2247 TEST_REQUIRES_ARM_NEON;
2248 for (uint32_t n = 16; n <= 24; n += 8) {
2249 for (size_t k = 1; k <= 80; k += 17) {
2250 GemmMicrokernelTester()
2251 .mr(2)
2252 .nr(8)
2253 .kr(2)
2254 .sr(1)
2255 .m(2)
2256 .n(n)
2257 .k(k)
2258 .ks(3)
2259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2260 }
2261 }
2262 }
2263
2264 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, strided_cm_subtile) {
2265 TEST_REQUIRES_ARM_NEON;
2266 for (size_t k = 1; k <= 80; k += 17) {
2267 for (uint32_t n = 1; n <= 8; n++) {
2268 for (uint32_t m = 1; m <= 2; m++) {
2269 GemmMicrokernelTester()
2270 .mr(2)
2271 .nr(8)
2272 .kr(2)
2273 .sr(1)
2274 .m(m)
2275 .n(n)
2276 .k(k)
2277 .cm_stride(11)
2278 .iterations(1)
2279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2280 }
2281 }
2282 }
2283 }
2284
2285 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, a_offset) {
2286 TEST_REQUIRES_ARM_NEON;
2287 for (size_t k = 1; k <= 80; k += 17) {
2288 GemmMicrokernelTester()
2289 .mr(2)
2290 .nr(8)
2291 .kr(2)
2292 .sr(1)
2293 .m(2)
2294 .n(8)
2295 .k(k)
2296 .ks(3)
2297 .a_offset(163)
2298 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2299 }
2300 }
2301
2302 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, zero) {
2303 TEST_REQUIRES_ARM_NEON;
2304 for (size_t k = 1; k <= 80; k += 17) {
2305 for (uint32_t mz = 0; mz < 2; mz++) {
2306 GemmMicrokernelTester()
2307 .mr(2)
2308 .nr(8)
2309 .kr(2)
2310 .sr(1)
2311 .m(2)
2312 .n(8)
2313 .k(k)
2314 .ks(3)
2315 .a_offset(163)
2316 .zero_index(mz)
2317 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2318 }
2319 }
2320 }
2321
2322 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, qmin) {
2323 TEST_REQUIRES_ARM_NEON;
2324 GemmMicrokernelTester()
2325 .mr(2)
2326 .nr(8)
2327 .kr(2)
2328 .sr(1)
2329 .m(2)
2330 .n(8)
2331 .k(16)
2332 .qmin(128)
2333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2334 }
2335
2336 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, qmax) {
2337 TEST_REQUIRES_ARM_NEON;
2338 GemmMicrokernelTester()
2339 .mr(2)
2340 .nr(8)
2341 .kr(2)
2342 .sr(1)
2343 .m(2)
2344 .n(8)
2345 .k(16)
2346 .qmax(128)
2347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2348 }
2349
2350 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2__NEON_MLAL_LD2R, strided_cm) {
2351 TEST_REQUIRES_ARM_NEON;
2352 GemmMicrokernelTester()
2353 .mr(2)
2354 .nr(8)
2355 .kr(2)
2356 .sr(1)
2357 .m(2)
2358 .n(8)
2359 .k(16)
2360 .cm_stride(11)
2361 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2362 }
2363#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
2364
2365
2366#if XNN_ARCH_ARM || XNN_ARCH_ARM64
2367 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_eq_8) {
2368 TEST_REQUIRES_ARM_NEON;
2369 GemmMicrokernelTester()
2370 .mr(3)
2371 .nr(8)
2372 .kr(2)
2373 .sr(1)
2374 .m(3)
2375 .n(8)
2376 .k(8)
2377 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2378 }
2379
2380 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, strided_cn) {
2381 TEST_REQUIRES_ARM_NEON;
2382 GemmMicrokernelTester()
2383 .mr(3)
2384 .nr(8)
2385 .kr(2)
2386 .sr(1)
2387 .m(3)
2388 .n(8)
2389 .k(8)
2390 .cn_stride(11)
2391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2392 }
2393
2394 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_eq_8_subtile) {
2395 TEST_REQUIRES_ARM_NEON;
2396 for (uint32_t n = 1; n <= 8; n++) {
2397 for (uint32_t m = 1; m <= 3; m++) {
2398 GemmMicrokernelTester()
2399 .mr(3)
2400 .nr(8)
2401 .kr(2)
2402 .sr(1)
2403 .m(m)
2404 .n(n)
2405 .k(8)
2406 .iterations(1)
2407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2408 }
2409 }
2410 }
2411
2412 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_eq_8_subtile_m) {
2413 TEST_REQUIRES_ARM_NEON;
2414 for (uint32_t m = 1; m <= 3; m++) {
2415 GemmMicrokernelTester()
2416 .mr(3)
2417 .nr(8)
2418 .kr(2)
2419 .sr(1)
2420 .m(m)
2421 .n(8)
2422 .k(8)
2423 .iterations(1)
2424 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2425 }
2426 }
2427
2428 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_eq_8_subtile_n) {
2429 TEST_REQUIRES_ARM_NEON;
2430 for (uint32_t n = 1; n <= 8; n++) {
2431 GemmMicrokernelTester()
2432 .mr(3)
2433 .nr(8)
2434 .kr(2)
2435 .sr(1)
2436 .m(3)
2437 .n(n)
2438 .k(8)
2439 .iterations(1)
2440 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2441 }
2442 }
2443
2444 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_lt_8) {
2445 TEST_REQUIRES_ARM_NEON;
2446 for (size_t k = 1; k < 8; k++) {
2447 GemmMicrokernelTester()
2448 .mr(3)
2449 .nr(8)
2450 .kr(2)
2451 .sr(1)
2452 .m(3)
2453 .n(8)
2454 .k(k)
2455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2456 }
2457 }
2458
2459 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_lt_8_subtile) {
2460 TEST_REQUIRES_ARM_NEON;
2461 for (size_t k = 1; k < 8; k++) {
2462 for (uint32_t n = 1; n <= 8; n++) {
2463 for (uint32_t m = 1; m <= 3; m++) {
2464 GemmMicrokernelTester()
2465 .mr(3)
2466 .nr(8)
2467 .kr(2)
2468 .sr(1)
2469 .m(m)
2470 .n(n)
2471 .k(k)
2472 .iterations(1)
2473 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2474 }
2475 }
2476 }
2477 }
2478
2479 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_gt_8) {
2480 TEST_REQUIRES_ARM_NEON;
2481 for (size_t k = 9; k < 16; k++) {
2482 GemmMicrokernelTester()
2483 .mr(3)
2484 .nr(8)
2485 .kr(2)
2486 .sr(1)
2487 .m(3)
2488 .n(8)
2489 .k(k)
2490 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2491 }
2492 }
2493
2494 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_gt_8_subtile) {
2495 TEST_REQUIRES_ARM_NEON;
2496 for (size_t k = 9; k < 16; k++) {
2497 for (uint32_t n = 1; n <= 8; n++) {
2498 for (uint32_t m = 1; m <= 3; m++) {
2499 GemmMicrokernelTester()
2500 .mr(3)
2501 .nr(8)
2502 .kr(2)
2503 .sr(1)
2504 .m(m)
2505 .n(n)
2506 .k(k)
2507 .iterations(1)
2508 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2509 }
2510 }
2511 }
2512 }
2513
2514 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_div_8) {
2515 TEST_REQUIRES_ARM_NEON;
2516 for (size_t k = 16; k <= 80; k += 8) {
2517 GemmMicrokernelTester()
2518 .mr(3)
2519 .nr(8)
2520 .kr(2)
2521 .sr(1)
2522 .m(3)
2523 .n(8)
2524 .k(k)
2525 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2526 }
2527 }
2528
2529 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, k_div_8_subtile) {
2530 TEST_REQUIRES_ARM_NEON;
2531 for (size_t k = 16; k <= 80; k += 8) {
2532 for (uint32_t n = 1; n <= 8; n++) {
2533 for (uint32_t m = 1; m <= 3; m++) {
2534 GemmMicrokernelTester()
2535 .mr(3)
2536 .nr(8)
2537 .kr(2)
2538 .sr(1)
2539 .m(m)
2540 .n(n)
2541 .k(k)
2542 .iterations(1)
2543 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2544 }
2545 }
2546 }
2547 }
2548
2549 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_gt_8) {
2550 TEST_REQUIRES_ARM_NEON;
2551 for (uint32_t n = 9; n < 16; n++) {
2552 for (size_t k = 1; k <= 40; k += 9) {
2553 GemmMicrokernelTester()
2554 .mr(3)
2555 .nr(8)
2556 .kr(2)
2557 .sr(1)
2558 .m(3)
2559 .n(n)
2560 .k(k)
2561 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2562 }
2563 }
2564 }
2565
2566 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_gt_8_strided_cn) {
2567 TEST_REQUIRES_ARM_NEON;
2568 for (uint32_t n = 9; n < 16; n++) {
2569 for (size_t k = 1; k <= 40; k += 9) {
2570 GemmMicrokernelTester()
2571 .mr(3)
2572 .nr(8)
2573 .kr(2)
2574 .sr(1)
2575 .m(3)
2576 .n(n)
2577 .k(k)
2578 .cn_stride(11)
2579 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2580 }
2581 }
2582 }
2583
2584 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_gt_8_subtile) {
2585 TEST_REQUIRES_ARM_NEON;
2586 for (uint32_t n = 9; n < 16; n++) {
2587 for (size_t k = 1; k <= 40; k += 9) {
2588 for (uint32_t m = 1; m <= 3; m++) {
2589 GemmMicrokernelTester()
2590 .mr(3)
2591 .nr(8)
2592 .kr(2)
2593 .sr(1)
2594 .m(m)
2595 .n(n)
2596 .k(k)
2597 .iterations(1)
2598 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2599 }
2600 }
2601 }
2602 }
2603
2604 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_div_8) {
2605 TEST_REQUIRES_ARM_NEON;
2606 for (uint32_t n = 16; n <= 24; n += 8) {
2607 for (size_t k = 1; k <= 40; k += 9) {
2608 GemmMicrokernelTester()
2609 .mr(3)
2610 .nr(8)
2611 .kr(2)
2612 .sr(1)
2613 .m(3)
2614 .n(n)
2615 .k(k)
2616 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2617 }
2618 }
2619 }
2620
2621 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_div_8_strided_cn) {
2622 TEST_REQUIRES_ARM_NEON;
2623 for (uint32_t n = 16; n <= 24; n += 8) {
2624 for (size_t k = 1; k <= 40; k += 9) {
2625 GemmMicrokernelTester()
2626 .mr(3)
2627 .nr(8)
2628 .kr(2)
2629 .sr(1)
2630 .m(3)
2631 .n(n)
2632 .k(k)
2633 .cn_stride(11)
2634 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2635 }
2636 }
2637 }
2638
2639 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_div_8_subtile) {
2640 TEST_REQUIRES_ARM_NEON;
2641 for (uint32_t n = 16; n <= 24; n += 8) {
2642 for (size_t k = 1; k <= 40; k += 9) {
2643 for (uint32_t m = 1; m <= 3; m++) {
2644 GemmMicrokernelTester()
2645 .mr(3)
2646 .nr(8)
2647 .kr(2)
2648 .sr(1)
2649 .m(m)
2650 .n(n)
2651 .k(k)
2652 .iterations(1)
2653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2654 }
2655 }
2656 }
2657 }
2658
2659 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, small_kernel) {
2660 TEST_REQUIRES_ARM_NEON;
2661 for (size_t k = 1; k <= 40; k += 9) {
2662 GemmMicrokernelTester()
2663 .mr(3)
2664 .nr(8)
2665 .kr(2)
2666 .sr(1)
2667 .m(3)
2668 .n(8)
2669 .k(k)
2670 .ks(3)
2671 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2672 }
2673 }
2674
2675 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, small_kernel_subtile) {
2676 TEST_REQUIRES_ARM_NEON;
2677 for (size_t k = 1; k <= 40; k += 9) {
2678 for (uint32_t n = 1; n <= 8; n++) {
2679 for (uint32_t m = 1; m <= 3; m++) {
2680 GemmMicrokernelTester()
2681 .mr(3)
2682 .nr(8)
2683 .kr(2)
2684 .sr(1)
2685 .m(m)
2686 .n(n)
2687 .k(k)
2688 .ks(3)
2689 .iterations(1)
2690 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2691 }
2692 }
2693 }
2694 }
2695
2696 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_gt_8_small_kernel) {
2697 TEST_REQUIRES_ARM_NEON;
2698 for (uint32_t n = 9; n < 16; n++) {
2699 for (size_t k = 1; k <= 40; k += 9) {
2700 GemmMicrokernelTester()
2701 .mr(3)
2702 .nr(8)
2703 .kr(2)
2704 .sr(1)
2705 .m(3)
2706 .n(n)
2707 .k(k)
2708 .ks(3)
2709 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2710 }
2711 }
2712 }
2713
2714 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, n_div_8_small_kernel) {
2715 TEST_REQUIRES_ARM_NEON;
2716 for (uint32_t n = 16; n <= 24; n += 8) {
2717 for (size_t k = 1; k <= 40; k += 9) {
2718 GemmMicrokernelTester()
2719 .mr(3)
2720 .nr(8)
2721 .kr(2)
2722 .sr(1)
2723 .m(3)
2724 .n(n)
2725 .k(k)
2726 .ks(3)
2727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2728 }
2729 }
2730 }
2731
2732 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, strided_cm_subtile) {
2733 TEST_REQUIRES_ARM_NEON;
2734 for (size_t k = 1; k <= 40; k += 9) {
2735 for (uint32_t n = 1; n <= 8; n++) {
2736 for (uint32_t m = 1; m <= 3; m++) {
2737 GemmMicrokernelTester()
2738 .mr(3)
2739 .nr(8)
2740 .kr(2)
2741 .sr(1)
2742 .m(m)
2743 .n(n)
2744 .k(k)
2745 .cm_stride(11)
2746 .iterations(1)
2747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2748 }
2749 }
2750 }
2751 }
2752
2753 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, a_offset) {
2754 TEST_REQUIRES_ARM_NEON;
2755 for (size_t k = 1; k <= 40; k += 9) {
2756 GemmMicrokernelTester()
2757 .mr(3)
2758 .nr(8)
2759 .kr(2)
2760 .sr(1)
2761 .m(3)
2762 .n(8)
2763 .k(k)
2764 .ks(3)
2765 .a_offset(127)
2766 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2767 }
2768 }
2769
2770 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, zero) {
2771 TEST_REQUIRES_ARM_NEON;
2772 for (size_t k = 1; k <= 40; k += 9) {
2773 for (uint32_t mz = 0; mz < 3; mz++) {
2774 GemmMicrokernelTester()
2775 .mr(3)
2776 .nr(8)
2777 .kr(2)
2778 .sr(1)
2779 .m(3)
2780 .n(8)
2781 .k(k)
2782 .ks(3)
2783 .a_offset(127)
2784 .zero_index(mz)
2785 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2786 }
2787 }
2788 }
2789
2790 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, qmin) {
2791 TEST_REQUIRES_ARM_NEON;
2792 GemmMicrokernelTester()
2793 .mr(3)
2794 .nr(8)
2795 .kr(2)
2796 .sr(1)
2797 .m(3)
2798 .n(8)
2799 .k(8)
2800 .qmin(128)
2801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2802 }
2803
2804 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, qmax) {
2805 TEST_REQUIRES_ARM_NEON;
2806 GemmMicrokernelTester()
2807 .mr(3)
2808 .nr(8)
2809 .kr(2)
2810 .sr(1)
2811 .m(3)
2812 .n(8)
2813 .k(8)
2814 .qmax(128)
2815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2816 }
2817
2818 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MULL_LD4R, strided_cm) {
2819 TEST_REQUIRES_ARM_NEON;
2820 GemmMicrokernelTester()
2821 .mr(3)
2822 .nr(8)
2823 .kr(2)
2824 .sr(1)
2825 .m(3)
2826 .n(8)
2827 .k(8)
2828 .cm_stride(11)
2829 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2830 }
2831#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
2832
2833
2834#if XNN_ARCH_ARM || XNN_ARCH_ARM64
2835 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_eq_8) {
2836 TEST_REQUIRES_ARM_NEON;
2837 GemmMicrokernelTester()
2838 .mr(4)
2839 .nr(8)
2840 .kr(2)
2841 .sr(1)
2842 .m(4)
2843 .n(8)
2844 .k(8)
2845 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2846 }
2847
2848 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, strided_cn) {
2849 TEST_REQUIRES_ARM_NEON;
2850 GemmMicrokernelTester()
2851 .mr(4)
2852 .nr(8)
2853 .kr(2)
2854 .sr(1)
2855 .m(4)
2856 .n(8)
2857 .k(8)
2858 .cn_stride(11)
2859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2860 }
2861
2862 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_eq_8_subtile) {
2863 TEST_REQUIRES_ARM_NEON;
2864 for (uint32_t n = 1; n <= 8; n++) {
2865 for (uint32_t m = 1; m <= 4; m++) {
2866 GemmMicrokernelTester()
2867 .mr(4)
2868 .nr(8)
2869 .kr(2)
2870 .sr(1)
2871 .m(m)
2872 .n(n)
2873 .k(8)
2874 .iterations(1)
2875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2876 }
2877 }
2878 }
2879
2880 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_eq_8_subtile_m) {
2881 TEST_REQUIRES_ARM_NEON;
2882 for (uint32_t m = 1; m <= 4; m++) {
2883 GemmMicrokernelTester()
2884 .mr(4)
2885 .nr(8)
2886 .kr(2)
2887 .sr(1)
2888 .m(m)
2889 .n(8)
2890 .k(8)
2891 .iterations(1)
2892 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2893 }
2894 }
2895
2896 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_eq_8_subtile_n) {
2897 TEST_REQUIRES_ARM_NEON;
2898 for (uint32_t n = 1; n <= 8; n++) {
2899 GemmMicrokernelTester()
2900 .mr(4)
2901 .nr(8)
2902 .kr(2)
2903 .sr(1)
2904 .m(4)
2905 .n(n)
2906 .k(8)
2907 .iterations(1)
2908 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2909 }
2910 }
2911
2912 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_lt_8) {
2913 TEST_REQUIRES_ARM_NEON;
2914 for (size_t k = 1; k < 8; k++) {
2915 GemmMicrokernelTester()
2916 .mr(4)
2917 .nr(8)
2918 .kr(2)
2919 .sr(1)
2920 .m(4)
2921 .n(8)
2922 .k(k)
2923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2924 }
2925 }
2926
2927 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_lt_8_subtile) {
2928 TEST_REQUIRES_ARM_NEON;
2929 for (size_t k = 1; k < 8; k++) {
2930 for (uint32_t n = 1; n <= 8; n++) {
2931 for (uint32_t m = 1; m <= 4; m++) {
2932 GemmMicrokernelTester()
2933 .mr(4)
2934 .nr(8)
2935 .kr(2)
2936 .sr(1)
2937 .m(m)
2938 .n(n)
2939 .k(k)
2940 .iterations(1)
2941 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2942 }
2943 }
2944 }
2945 }
2946
2947 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_gt_8) {
2948 TEST_REQUIRES_ARM_NEON;
2949 for (size_t k = 9; k < 16; k++) {
2950 GemmMicrokernelTester()
2951 .mr(4)
2952 .nr(8)
2953 .kr(2)
2954 .sr(1)
2955 .m(4)
2956 .n(8)
2957 .k(k)
2958 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2959 }
2960 }
2961
2962 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_gt_8_subtile) {
2963 TEST_REQUIRES_ARM_NEON;
2964 for (size_t k = 9; k < 16; k++) {
2965 for (uint32_t n = 1; n <= 8; n++) {
2966 for (uint32_t m = 1; m <= 4; m++) {
2967 GemmMicrokernelTester()
2968 .mr(4)
2969 .nr(8)
2970 .kr(2)
2971 .sr(1)
2972 .m(m)
2973 .n(n)
2974 .k(k)
2975 .iterations(1)
2976 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2977 }
2978 }
2979 }
2980 }
2981
2982 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_div_8) {
2983 TEST_REQUIRES_ARM_NEON;
2984 for (size_t k = 16; k <= 80; k += 8) {
2985 GemmMicrokernelTester()
2986 .mr(4)
2987 .nr(8)
2988 .kr(2)
2989 .sr(1)
2990 .m(4)
2991 .n(8)
2992 .k(k)
2993 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
2994 }
2995 }
2996
2997 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, k_div_8_subtile) {
2998 TEST_REQUIRES_ARM_NEON;
2999 for (size_t k = 16; k <= 80; k += 8) {
3000 for (uint32_t n = 1; n <= 8; n++) {
3001 for (uint32_t m = 1; m <= 4; m++) {
3002 GemmMicrokernelTester()
3003 .mr(4)
3004 .nr(8)
3005 .kr(2)
3006 .sr(1)
3007 .m(m)
3008 .n(n)
3009 .k(k)
3010 .iterations(1)
3011 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3012 }
3013 }
3014 }
3015 }
3016
3017 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_gt_8) {
3018 TEST_REQUIRES_ARM_NEON;
3019 for (uint32_t n = 9; n < 16; n++) {
3020 for (size_t k = 1; k <= 40; k += 9) {
3021 GemmMicrokernelTester()
3022 .mr(4)
3023 .nr(8)
3024 .kr(2)
3025 .sr(1)
3026 .m(4)
3027 .n(n)
3028 .k(k)
3029 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3030 }
3031 }
3032 }
3033
3034 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_gt_8_strided_cn) {
3035 TEST_REQUIRES_ARM_NEON;
3036 for (uint32_t n = 9; n < 16; n++) {
3037 for (size_t k = 1; k <= 40; k += 9) {
3038 GemmMicrokernelTester()
3039 .mr(4)
3040 .nr(8)
3041 .kr(2)
3042 .sr(1)
3043 .m(4)
3044 .n(n)
3045 .k(k)
3046 .cn_stride(11)
3047 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3048 }
3049 }
3050 }
3051
3052 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_gt_8_subtile) {
3053 TEST_REQUIRES_ARM_NEON;
3054 for (uint32_t n = 9; n < 16; n++) {
3055 for (size_t k = 1; k <= 40; k += 9) {
3056 for (uint32_t m = 1; m <= 4; m++) {
3057 GemmMicrokernelTester()
3058 .mr(4)
3059 .nr(8)
3060 .kr(2)
3061 .sr(1)
3062 .m(m)
3063 .n(n)
3064 .k(k)
3065 .iterations(1)
3066 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3067 }
3068 }
3069 }
3070 }
3071
3072 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_div_8) {
3073 TEST_REQUIRES_ARM_NEON;
3074 for (uint32_t n = 16; n <= 24; n += 8) {
3075 for (size_t k = 1; k <= 40; k += 9) {
3076 GemmMicrokernelTester()
3077 .mr(4)
3078 .nr(8)
3079 .kr(2)
3080 .sr(1)
3081 .m(4)
3082 .n(n)
3083 .k(k)
3084 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3085 }
3086 }
3087 }
3088
3089 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_div_8_strided_cn) {
3090 TEST_REQUIRES_ARM_NEON;
3091 for (uint32_t n = 16; n <= 24; n += 8) {
3092 for (size_t k = 1; k <= 40; k += 9) {
3093 GemmMicrokernelTester()
3094 .mr(4)
3095 .nr(8)
3096 .kr(2)
3097 .sr(1)
3098 .m(4)
3099 .n(n)
3100 .k(k)
3101 .cn_stride(11)
3102 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3103 }
3104 }
3105 }
3106
3107 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_div_8_subtile) {
3108 TEST_REQUIRES_ARM_NEON;
3109 for (uint32_t n = 16; n <= 24; n += 8) {
3110 for (size_t k = 1; k <= 40; k += 9) {
3111 for (uint32_t m = 1; m <= 4; m++) {
3112 GemmMicrokernelTester()
3113 .mr(4)
3114 .nr(8)
3115 .kr(2)
3116 .sr(1)
3117 .m(m)
3118 .n(n)
3119 .k(k)
3120 .iterations(1)
3121 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3122 }
3123 }
3124 }
3125 }
3126
3127 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, small_kernel) {
3128 TEST_REQUIRES_ARM_NEON;
3129 for (size_t k = 1; k <= 40; k += 9) {
3130 GemmMicrokernelTester()
3131 .mr(4)
3132 .nr(8)
3133 .kr(2)
3134 .sr(1)
3135 .m(4)
3136 .n(8)
3137 .k(k)
3138 .ks(3)
3139 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3140 }
3141 }
3142
3143 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, small_kernel_subtile) {
3144 TEST_REQUIRES_ARM_NEON;
3145 for (size_t k = 1; k <= 40; k += 9) {
3146 for (uint32_t n = 1; n <= 8; n++) {
3147 for (uint32_t m = 1; m <= 4; m++) {
3148 GemmMicrokernelTester()
3149 .mr(4)
3150 .nr(8)
3151 .kr(2)
3152 .sr(1)
3153 .m(m)
3154 .n(n)
3155 .k(k)
3156 .ks(3)
3157 .iterations(1)
3158 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3159 }
3160 }
3161 }
3162 }
3163
3164 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_gt_8_small_kernel) {
3165 TEST_REQUIRES_ARM_NEON;
3166 for (uint32_t n = 9; n < 16; n++) {
3167 for (size_t k = 1; k <= 40; k += 9) {
3168 GemmMicrokernelTester()
3169 .mr(4)
3170 .nr(8)
3171 .kr(2)
3172 .sr(1)
3173 .m(4)
3174 .n(n)
3175 .k(k)
3176 .ks(3)
3177 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3178 }
3179 }
3180 }
3181
3182 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, n_div_8_small_kernel) {
3183 TEST_REQUIRES_ARM_NEON;
3184 for (uint32_t n = 16; n <= 24; n += 8) {
3185 for (size_t k = 1; k <= 40; k += 9) {
3186 GemmMicrokernelTester()
3187 .mr(4)
3188 .nr(8)
3189 .kr(2)
3190 .sr(1)
3191 .m(4)
3192 .n(n)
3193 .k(k)
3194 .ks(3)
3195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3196 }
3197 }
3198 }
3199
3200 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, strided_cm_subtile) {
3201 TEST_REQUIRES_ARM_NEON;
3202 for (size_t k = 1; k <= 40; k += 9) {
3203 for (uint32_t n = 1; n <= 8; n++) {
3204 for (uint32_t m = 1; m <= 4; m++) {
3205 GemmMicrokernelTester()
3206 .mr(4)
3207 .nr(8)
3208 .kr(2)
3209 .sr(1)
3210 .m(m)
3211 .n(n)
3212 .k(k)
3213 .cm_stride(11)
3214 .iterations(1)
3215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3216 }
3217 }
3218 }
3219 }
3220
3221 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, a_offset) {
3222 TEST_REQUIRES_ARM_NEON;
3223 for (size_t k = 1; k <= 40; k += 9) {
3224 GemmMicrokernelTester()
3225 .mr(4)
3226 .nr(8)
3227 .kr(2)
3228 .sr(1)
3229 .m(4)
3230 .n(8)
3231 .k(k)
3232 .ks(3)
3233 .a_offset(163)
3234 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3235 }
3236 }
3237
3238 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, zero) {
3239 TEST_REQUIRES_ARM_NEON;
3240 for (size_t k = 1; k <= 40; k += 9) {
3241 for (uint32_t mz = 0; mz < 4; mz++) {
3242 GemmMicrokernelTester()
3243 .mr(4)
3244 .nr(8)
3245 .kr(2)
3246 .sr(1)
3247 .m(4)
3248 .n(8)
3249 .k(k)
3250 .ks(3)
3251 .a_offset(163)
3252 .zero_index(mz)
3253 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3254 }
3255 }
3256 }
3257
3258 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, qmin) {
3259 TEST_REQUIRES_ARM_NEON;
3260 GemmMicrokernelTester()
3261 .mr(4)
3262 .nr(8)
3263 .kr(2)
3264 .sr(1)
3265 .m(4)
3266 .n(8)
3267 .k(8)
3268 .qmin(128)
3269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3270 }
3271
3272 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, qmax) {
3273 TEST_REQUIRES_ARM_NEON;
3274 GemmMicrokernelTester()
3275 .mr(4)
3276 .nr(8)
3277 .kr(2)
3278 .sr(1)
3279 .m(4)
3280 .n(8)
3281 .k(8)
3282 .qmax(128)
3283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3284 }
3285
3286 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_LD4R, strided_cm) {
3287 TEST_REQUIRES_ARM_NEON;
3288 GemmMicrokernelTester()
3289 .mr(4)
3290 .nr(8)
3291 .kr(2)
3292 .sr(1)
3293 .m(4)
3294 .n(8)
3295 .k(8)
3296 .cm_stride(11)
3297 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3298 }
3299#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
3300
3301
3302#if XNN_ARCH_ARM || XNN_ARCH_ARM64
3303 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_eq_16) {
3304 TEST_REQUIRES_ARM_NEON;
3305 GemmMicrokernelTester()
3306 .mr(3)
3307 .nr(8)
3308 .kr(2)
3309 .sr(1)
3310 .m(3)
3311 .n(8)
3312 .k(16)
3313 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3314 }
3315
3316 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, strided_cn) {
3317 TEST_REQUIRES_ARM_NEON;
3318 GemmMicrokernelTester()
3319 .mr(3)
3320 .nr(8)
3321 .kr(2)
3322 .sr(1)
3323 .m(3)
3324 .n(8)
3325 .k(16)
3326 .cn_stride(11)
3327 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3328 }
3329
3330 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_eq_16_subtile) {
3331 TEST_REQUIRES_ARM_NEON;
3332 for (uint32_t n = 1; n <= 8; n++) {
3333 for (uint32_t m = 1; m <= 3; m++) {
3334 GemmMicrokernelTester()
3335 .mr(3)
3336 .nr(8)
3337 .kr(2)
3338 .sr(1)
3339 .m(m)
3340 .n(n)
3341 .k(16)
3342 .iterations(1)
3343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3344 }
3345 }
3346 }
3347
3348 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_eq_16_subtile_m) {
3349 TEST_REQUIRES_ARM_NEON;
3350 for (uint32_t m = 1; m <= 3; m++) {
3351 GemmMicrokernelTester()
3352 .mr(3)
3353 .nr(8)
3354 .kr(2)
3355 .sr(1)
3356 .m(m)
3357 .n(8)
3358 .k(16)
3359 .iterations(1)
3360 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3361 }
3362 }
3363
3364 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_eq_16_subtile_n) {
3365 TEST_REQUIRES_ARM_NEON;
3366 for (uint32_t n = 1; n <= 8; n++) {
3367 GemmMicrokernelTester()
3368 .mr(3)
3369 .nr(8)
3370 .kr(2)
3371 .sr(1)
3372 .m(3)
3373 .n(n)
3374 .k(16)
3375 .iterations(1)
3376 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3377 }
3378 }
3379
3380 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_lt_16) {
3381 TEST_REQUIRES_ARM_NEON;
3382 for (size_t k = 1; k < 16; k++) {
3383 GemmMicrokernelTester()
3384 .mr(3)
3385 .nr(8)
3386 .kr(2)
3387 .sr(1)
3388 .m(3)
3389 .n(8)
3390 .k(k)
3391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3392 }
3393 }
3394
3395 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_lt_16_subtile) {
3396 TEST_REQUIRES_ARM_NEON;
3397 for (size_t k = 1; k < 16; k++) {
3398 for (uint32_t n = 1; n <= 8; n++) {
3399 for (uint32_t m = 1; m <= 3; m++) {
3400 GemmMicrokernelTester()
3401 .mr(3)
3402 .nr(8)
3403 .kr(2)
3404 .sr(1)
3405 .m(m)
3406 .n(n)
3407 .k(k)
3408 .iterations(1)
3409 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3410 }
3411 }
3412 }
3413 }
3414
3415 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_gt_16) {
3416 TEST_REQUIRES_ARM_NEON;
3417 for (size_t k = 17; k < 32; k++) {
3418 GemmMicrokernelTester()
3419 .mr(3)
3420 .nr(8)
3421 .kr(2)
3422 .sr(1)
3423 .m(3)
3424 .n(8)
3425 .k(k)
3426 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3427 }
3428 }
3429
3430 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_gt_16_subtile) {
3431 TEST_REQUIRES_ARM_NEON;
3432 for (size_t k = 17; k < 32; k++) {
3433 for (uint32_t n = 1; n <= 8; n++) {
3434 for (uint32_t m = 1; m <= 3; m++) {
3435 GemmMicrokernelTester()
3436 .mr(3)
3437 .nr(8)
3438 .kr(2)
3439 .sr(1)
3440 .m(m)
3441 .n(n)
3442 .k(k)
3443 .iterations(1)
3444 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3445 }
3446 }
3447 }
3448 }
3449
3450 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_div_16) {
3451 TEST_REQUIRES_ARM_NEON;
3452 for (size_t k = 32; k <= 160; k += 16) {
3453 GemmMicrokernelTester()
3454 .mr(3)
3455 .nr(8)
3456 .kr(2)
3457 .sr(1)
3458 .m(3)
3459 .n(8)
3460 .k(k)
3461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3462 }
3463 }
3464
3465 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, k_div_16_subtile) {
3466 TEST_REQUIRES_ARM_NEON;
3467 for (size_t k = 32; k <= 160; k += 16) {
3468 for (uint32_t n = 1; n <= 8; n++) {
3469 for (uint32_t m = 1; m <= 3; m++) {
3470 GemmMicrokernelTester()
3471 .mr(3)
3472 .nr(8)
3473 .kr(2)
3474 .sr(1)
3475 .m(m)
3476 .n(n)
3477 .k(k)
3478 .iterations(1)
3479 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3480 }
3481 }
3482 }
3483 }
3484
3485 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_gt_8) {
3486 TEST_REQUIRES_ARM_NEON;
3487 for (uint32_t n = 9; n < 16; n++) {
3488 for (size_t k = 1; k <= 80; k += 17) {
3489 GemmMicrokernelTester()
3490 .mr(3)
3491 .nr(8)
3492 .kr(2)
3493 .sr(1)
3494 .m(3)
3495 .n(n)
3496 .k(k)
3497 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3498 }
3499 }
3500 }
3501
3502 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_gt_8_strided_cn) {
3503 TEST_REQUIRES_ARM_NEON;
3504 for (uint32_t n = 9; n < 16; n++) {
3505 for (size_t k = 1; k <= 80; k += 17) {
3506 GemmMicrokernelTester()
3507 .mr(3)
3508 .nr(8)
3509 .kr(2)
3510 .sr(1)
3511 .m(3)
3512 .n(n)
3513 .k(k)
3514 .cn_stride(11)
3515 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3516 }
3517 }
3518 }
3519
3520 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_gt_8_subtile) {
3521 TEST_REQUIRES_ARM_NEON;
3522 for (uint32_t n = 9; n < 16; n++) {
3523 for (size_t k = 1; k <= 80; k += 17) {
3524 for (uint32_t m = 1; m <= 3; m++) {
3525 GemmMicrokernelTester()
3526 .mr(3)
3527 .nr(8)
3528 .kr(2)
3529 .sr(1)
3530 .m(m)
3531 .n(n)
3532 .k(k)
3533 .iterations(1)
3534 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3535 }
3536 }
3537 }
3538 }
3539
3540 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_div_8) {
3541 TEST_REQUIRES_ARM_NEON;
3542 for (uint32_t n = 16; n <= 24; n += 8) {
3543 for (size_t k = 1; k <= 80; k += 17) {
3544 GemmMicrokernelTester()
3545 .mr(3)
3546 .nr(8)
3547 .kr(2)
3548 .sr(1)
3549 .m(3)
3550 .n(n)
3551 .k(k)
3552 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3553 }
3554 }
3555 }
3556
3557 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_div_8_strided_cn) {
3558 TEST_REQUIRES_ARM_NEON;
3559 for (uint32_t n = 16; n <= 24; n += 8) {
3560 for (size_t k = 1; k <= 80; k += 17) {
3561 GemmMicrokernelTester()
3562 .mr(3)
3563 .nr(8)
3564 .kr(2)
3565 .sr(1)
3566 .m(3)
3567 .n(n)
3568 .k(k)
3569 .cn_stride(11)
3570 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3571 }
3572 }
3573 }
3574
3575 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_div_8_subtile) {
3576 TEST_REQUIRES_ARM_NEON;
3577 for (uint32_t n = 16; n <= 24; n += 8) {
3578 for (size_t k = 1; k <= 80; k += 17) {
3579 for (uint32_t m = 1; m <= 3; m++) {
3580 GemmMicrokernelTester()
3581 .mr(3)
3582 .nr(8)
3583 .kr(2)
3584 .sr(1)
3585 .m(m)
3586 .n(n)
3587 .k(k)
3588 .iterations(1)
3589 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3590 }
3591 }
3592 }
3593 }
3594
3595 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, small_kernel) {
3596 TEST_REQUIRES_ARM_NEON;
3597 for (size_t k = 1; k <= 80; k += 17) {
3598 GemmMicrokernelTester()
3599 .mr(3)
3600 .nr(8)
3601 .kr(2)
3602 .sr(1)
3603 .m(3)
3604 .n(8)
3605 .k(k)
3606 .ks(3)
3607 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3608 }
3609 }
3610
3611 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, small_kernel_subtile) {
3612 TEST_REQUIRES_ARM_NEON;
3613 for (size_t k = 1; k <= 80; k += 17) {
3614 for (uint32_t n = 1; n <= 8; n++) {
3615 for (uint32_t m = 1; m <= 3; m++) {
3616 GemmMicrokernelTester()
3617 .mr(3)
3618 .nr(8)
3619 .kr(2)
3620 .sr(1)
3621 .m(m)
3622 .n(n)
3623 .k(k)
3624 .ks(3)
3625 .iterations(1)
3626 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3627 }
3628 }
3629 }
3630 }
3631
3632 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_gt_8_small_kernel) {
3633 TEST_REQUIRES_ARM_NEON;
3634 for (uint32_t n = 9; n < 16; n++) {
3635 for (size_t k = 1; k <= 80; k += 17) {
3636 GemmMicrokernelTester()
3637 .mr(3)
3638 .nr(8)
3639 .kr(2)
3640 .sr(1)
3641 .m(3)
3642 .n(n)
3643 .k(k)
3644 .ks(3)
3645 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3646 }
3647 }
3648 }
3649
3650 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, n_div_8_small_kernel) {
3651 TEST_REQUIRES_ARM_NEON;
3652 for (uint32_t n = 16; n <= 24; n += 8) {
3653 for (size_t k = 1; k <= 80; k += 17) {
3654 GemmMicrokernelTester()
3655 .mr(3)
3656 .nr(8)
3657 .kr(2)
3658 .sr(1)
3659 .m(3)
3660 .n(n)
3661 .k(k)
3662 .ks(3)
3663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3664 }
3665 }
3666 }
3667
3668 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, strided_cm_subtile) {
3669 TEST_REQUIRES_ARM_NEON;
3670 for (size_t k = 1; k <= 80; k += 17) {
3671 for (uint32_t n = 1; n <= 8; n++) {
3672 for (uint32_t m = 1; m <= 3; m++) {
3673 GemmMicrokernelTester()
3674 .mr(3)
3675 .nr(8)
3676 .kr(2)
3677 .sr(1)
3678 .m(m)
3679 .n(n)
3680 .k(k)
3681 .cm_stride(11)
3682 .iterations(1)
3683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3684 }
3685 }
3686 }
3687 }
3688
3689 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, a_offset) {
3690 TEST_REQUIRES_ARM_NEON;
3691 for (size_t k = 1; k <= 80; k += 17) {
3692 GemmMicrokernelTester()
3693 .mr(3)
3694 .nr(8)
3695 .kr(2)
3696 .sr(1)
3697 .m(3)
3698 .n(8)
3699 .k(k)
3700 .ks(3)
3701 .a_offset(251)
3702 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3703 }
3704 }
3705
3706 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, zero) {
3707 TEST_REQUIRES_ARM_NEON;
3708 for (size_t k = 1; k <= 80; k += 17) {
3709 for (uint32_t mz = 0; mz < 3; mz++) {
3710 GemmMicrokernelTester()
3711 .mr(3)
3712 .nr(8)
3713 .kr(2)
3714 .sr(1)
3715 .m(3)
3716 .n(8)
3717 .k(k)
3718 .ks(3)
3719 .a_offset(251)
3720 .zero_index(mz)
3721 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3722 }
3723 }
3724 }
3725
3726 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, qmin) {
3727 TEST_REQUIRES_ARM_NEON;
3728 GemmMicrokernelTester()
3729 .mr(3)
3730 .nr(8)
3731 .kr(2)
3732 .sr(1)
3733 .m(3)
3734 .n(8)
3735 .k(16)
3736 .qmin(128)
3737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3738 }
3739
3740 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, qmax) {
3741 TEST_REQUIRES_ARM_NEON;
3742 GemmMicrokernelTester()
3743 .mr(3)
3744 .nr(8)
3745 .kr(2)
3746 .sr(1)
3747 .m(3)
3748 .n(8)
3749 .k(16)
3750 .qmax(128)
3751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3752 }
3753
3754 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_LD4R, strided_cm) {
3755 TEST_REQUIRES_ARM_NEON;
3756 GemmMicrokernelTester()
3757 .mr(3)
3758 .nr(8)
3759 .kr(2)
3760 .sr(1)
3761 .m(3)
3762 .n(8)
3763 .k(16)
3764 .cm_stride(11)
3765 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3766 }
3767#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
3768
3769
3770#if XNN_ARCH_ARM || XNN_ARCH_ARM64
3771 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_eq_16) {
3772 TEST_REQUIRES_ARM_NEON;
3773 GemmMicrokernelTester()
3774 .mr(4)
3775 .nr(8)
3776 .kr(2)
3777 .sr(1)
3778 .m(4)
3779 .n(8)
3780 .k(16)
3781 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3782 }
3783
3784 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, strided_cn) {
3785 TEST_REQUIRES_ARM_NEON;
3786 GemmMicrokernelTester()
3787 .mr(4)
3788 .nr(8)
3789 .kr(2)
3790 .sr(1)
3791 .m(4)
3792 .n(8)
3793 .k(16)
3794 .cn_stride(11)
3795 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3796 }
3797
3798 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_eq_16_subtile) {
3799 TEST_REQUIRES_ARM_NEON;
3800 for (uint32_t n = 1; n <= 8; n++) {
3801 for (uint32_t m = 1; m <= 4; m++) {
3802 GemmMicrokernelTester()
3803 .mr(4)
3804 .nr(8)
3805 .kr(2)
3806 .sr(1)
3807 .m(m)
3808 .n(n)
3809 .k(16)
3810 .iterations(1)
3811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3812 }
3813 }
3814 }
3815
3816 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_eq_16_subtile_m) {
3817 TEST_REQUIRES_ARM_NEON;
3818 for (uint32_t m = 1; m <= 4; m++) {
3819 GemmMicrokernelTester()
3820 .mr(4)
3821 .nr(8)
3822 .kr(2)
3823 .sr(1)
3824 .m(m)
3825 .n(8)
3826 .k(16)
3827 .iterations(1)
3828 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3829 }
3830 }
3831
3832 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_eq_16_subtile_n) {
3833 TEST_REQUIRES_ARM_NEON;
3834 for (uint32_t n = 1; n <= 8; n++) {
3835 GemmMicrokernelTester()
3836 .mr(4)
3837 .nr(8)
3838 .kr(2)
3839 .sr(1)
3840 .m(4)
3841 .n(n)
3842 .k(16)
3843 .iterations(1)
3844 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3845 }
3846 }
3847
3848 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_lt_16) {
3849 TEST_REQUIRES_ARM_NEON;
3850 for (size_t k = 1; k < 16; k++) {
3851 GemmMicrokernelTester()
3852 .mr(4)
3853 .nr(8)
3854 .kr(2)
3855 .sr(1)
3856 .m(4)
3857 .n(8)
3858 .k(k)
3859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3860 }
3861 }
3862
3863 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_lt_16_subtile) {
3864 TEST_REQUIRES_ARM_NEON;
3865 for (size_t k = 1; k < 16; k++) {
3866 for (uint32_t n = 1; n <= 8; n++) {
3867 for (uint32_t m = 1; m <= 4; m++) {
3868 GemmMicrokernelTester()
3869 .mr(4)
3870 .nr(8)
3871 .kr(2)
3872 .sr(1)
3873 .m(m)
3874 .n(n)
3875 .k(k)
3876 .iterations(1)
3877 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3878 }
3879 }
3880 }
3881 }
3882
3883 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_gt_16) {
3884 TEST_REQUIRES_ARM_NEON;
3885 for (size_t k = 17; k < 32; k++) {
3886 GemmMicrokernelTester()
3887 .mr(4)
3888 .nr(8)
3889 .kr(2)
3890 .sr(1)
3891 .m(4)
3892 .n(8)
3893 .k(k)
3894 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3895 }
3896 }
3897
3898 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_gt_16_subtile) {
3899 TEST_REQUIRES_ARM_NEON;
3900 for (size_t k = 17; k < 32; k++) {
3901 for (uint32_t n = 1; n <= 8; n++) {
3902 for (uint32_t m = 1; m <= 4; m++) {
3903 GemmMicrokernelTester()
3904 .mr(4)
3905 .nr(8)
3906 .kr(2)
3907 .sr(1)
3908 .m(m)
3909 .n(n)
3910 .k(k)
3911 .iterations(1)
3912 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3913 }
3914 }
3915 }
3916 }
3917
3918 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_div_16) {
3919 TEST_REQUIRES_ARM_NEON;
3920 for (size_t k = 32; k <= 160; k += 16) {
3921 GemmMicrokernelTester()
3922 .mr(4)
3923 .nr(8)
3924 .kr(2)
3925 .sr(1)
3926 .m(4)
3927 .n(8)
3928 .k(k)
3929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3930 }
3931 }
3932
3933 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, k_div_16_subtile) {
3934 TEST_REQUIRES_ARM_NEON;
3935 for (size_t k = 32; k <= 160; k += 16) {
3936 for (uint32_t n = 1; n <= 8; n++) {
3937 for (uint32_t m = 1; m <= 4; m++) {
3938 GemmMicrokernelTester()
3939 .mr(4)
3940 .nr(8)
3941 .kr(2)
3942 .sr(1)
3943 .m(m)
3944 .n(n)
3945 .k(k)
3946 .iterations(1)
3947 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3948 }
3949 }
3950 }
3951 }
3952
3953 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_gt_8) {
3954 TEST_REQUIRES_ARM_NEON;
3955 for (uint32_t n = 9; n < 16; n++) {
3956 for (size_t k = 1; k <= 80; k += 17) {
3957 GemmMicrokernelTester()
3958 .mr(4)
3959 .nr(8)
3960 .kr(2)
3961 .sr(1)
3962 .m(4)
3963 .n(n)
3964 .k(k)
3965 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3966 }
3967 }
3968 }
3969
3970 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_gt_8_strided_cn) {
3971 TEST_REQUIRES_ARM_NEON;
3972 for (uint32_t n = 9; n < 16; n++) {
3973 for (size_t k = 1; k <= 80; k += 17) {
3974 GemmMicrokernelTester()
3975 .mr(4)
3976 .nr(8)
3977 .kr(2)
3978 .sr(1)
3979 .m(4)
3980 .n(n)
3981 .k(k)
3982 .cn_stride(11)
3983 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
3984 }
3985 }
3986 }
3987
3988 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_gt_8_subtile) {
3989 TEST_REQUIRES_ARM_NEON;
3990 for (uint32_t n = 9; n < 16; n++) {
3991 for (size_t k = 1; k <= 80; k += 17) {
3992 for (uint32_t m = 1; m <= 4; m++) {
3993 GemmMicrokernelTester()
3994 .mr(4)
3995 .nr(8)
3996 .kr(2)
3997 .sr(1)
3998 .m(m)
3999 .n(n)
4000 .k(k)
4001 .iterations(1)
4002 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4003 }
4004 }
4005 }
4006 }
4007
4008 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_div_8) {
4009 TEST_REQUIRES_ARM_NEON;
4010 for (uint32_t n = 16; n <= 24; n += 8) {
4011 for (size_t k = 1; k <= 80; k += 17) {
4012 GemmMicrokernelTester()
4013 .mr(4)
4014 .nr(8)
4015 .kr(2)
4016 .sr(1)
4017 .m(4)
4018 .n(n)
4019 .k(k)
4020 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4021 }
4022 }
4023 }
4024
4025 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_div_8_strided_cn) {
4026 TEST_REQUIRES_ARM_NEON;
4027 for (uint32_t n = 16; n <= 24; n += 8) {
4028 for (size_t k = 1; k <= 80; k += 17) {
4029 GemmMicrokernelTester()
4030 .mr(4)
4031 .nr(8)
4032 .kr(2)
4033 .sr(1)
4034 .m(4)
4035 .n(n)
4036 .k(k)
4037 .cn_stride(11)
4038 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4039 }
4040 }
4041 }
4042
4043 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_div_8_subtile) {
4044 TEST_REQUIRES_ARM_NEON;
4045 for (uint32_t n = 16; n <= 24; n += 8) {
4046 for (size_t k = 1; k <= 80; k += 17) {
4047 for (uint32_t m = 1; m <= 4; m++) {
4048 GemmMicrokernelTester()
4049 .mr(4)
4050 .nr(8)
4051 .kr(2)
4052 .sr(1)
4053 .m(m)
4054 .n(n)
4055 .k(k)
4056 .iterations(1)
4057 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4058 }
4059 }
4060 }
4061 }
4062
4063 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, small_kernel) {
4064 TEST_REQUIRES_ARM_NEON;
4065 for (size_t k = 1; k <= 80; k += 17) {
4066 GemmMicrokernelTester()
4067 .mr(4)
4068 .nr(8)
4069 .kr(2)
4070 .sr(1)
4071 .m(4)
4072 .n(8)
4073 .k(k)
4074 .ks(3)
4075 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4076 }
4077 }
4078
4079 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, small_kernel_subtile) {
4080 TEST_REQUIRES_ARM_NEON;
4081 for (size_t k = 1; k <= 80; k += 17) {
4082 for (uint32_t n = 1; n <= 8; n++) {
4083 for (uint32_t m = 1; m <= 4; m++) {
4084 GemmMicrokernelTester()
4085 .mr(4)
4086 .nr(8)
4087 .kr(2)
4088 .sr(1)
4089 .m(m)
4090 .n(n)
4091 .k(k)
4092 .ks(3)
4093 .iterations(1)
4094 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4095 }
4096 }
4097 }
4098 }
4099
4100 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_gt_8_small_kernel) {
4101 TEST_REQUIRES_ARM_NEON;
4102 for (uint32_t n = 9; n < 16; n++) {
4103 for (size_t k = 1; k <= 80; k += 17) {
4104 GemmMicrokernelTester()
4105 .mr(4)
4106 .nr(8)
4107 .kr(2)
4108 .sr(1)
4109 .m(4)
4110 .n(n)
4111 .k(k)
4112 .ks(3)
4113 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4114 }
4115 }
4116 }
4117
4118 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, n_div_8_small_kernel) {
4119 TEST_REQUIRES_ARM_NEON;
4120 for (uint32_t n = 16; n <= 24; n += 8) {
4121 for (size_t k = 1; k <= 80; k += 17) {
4122 GemmMicrokernelTester()
4123 .mr(4)
4124 .nr(8)
4125 .kr(2)
4126 .sr(1)
4127 .m(4)
4128 .n(n)
4129 .k(k)
4130 .ks(3)
4131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4132 }
4133 }
4134 }
4135
4136 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, strided_cm_subtile) {
4137 TEST_REQUIRES_ARM_NEON;
4138 for (size_t k = 1; k <= 80; k += 17) {
4139 for (uint32_t n = 1; n <= 8; n++) {
4140 for (uint32_t m = 1; m <= 4; m++) {
4141 GemmMicrokernelTester()
4142 .mr(4)
4143 .nr(8)
4144 .kr(2)
4145 .sr(1)
4146 .m(m)
4147 .n(n)
4148 .k(k)
4149 .cm_stride(11)
4150 .iterations(1)
4151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4152 }
4153 }
4154 }
4155 }
4156
4157 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, a_offset) {
4158 TEST_REQUIRES_ARM_NEON;
4159 for (size_t k = 1; k <= 80; k += 17) {
4160 GemmMicrokernelTester()
4161 .mr(4)
4162 .nr(8)
4163 .kr(2)
4164 .sr(1)
4165 .m(4)
4166 .n(8)
4167 .k(k)
4168 .ks(3)
4169 .a_offset(331)
4170 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4171 }
4172 }
4173
4174 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, zero) {
4175 TEST_REQUIRES_ARM_NEON;
4176 for (size_t k = 1; k <= 80; k += 17) {
4177 for (uint32_t mz = 0; mz < 4; mz++) {
4178 GemmMicrokernelTester()
4179 .mr(4)
4180 .nr(8)
4181 .kr(2)
4182 .sr(1)
4183 .m(4)
4184 .n(8)
4185 .k(k)
4186 .ks(3)
4187 .a_offset(331)
4188 .zero_index(mz)
4189 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4190 }
4191 }
4192 }
4193
4194 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, qmin) {
4195 TEST_REQUIRES_ARM_NEON;
4196 GemmMicrokernelTester()
4197 .mr(4)
4198 .nr(8)
4199 .kr(2)
4200 .sr(1)
4201 .m(4)
4202 .n(8)
4203 .k(16)
4204 .qmin(128)
4205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4206 }
4207
4208 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, qmax) {
4209 TEST_REQUIRES_ARM_NEON;
4210 GemmMicrokernelTester()
4211 .mr(4)
4212 .nr(8)
4213 .kr(2)
4214 .sr(1)
4215 .m(4)
4216 .n(8)
4217 .k(16)
4218 .qmax(128)
4219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4220 }
4221
4222 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MLAL_LD4R, strided_cm) {
4223 TEST_REQUIRES_ARM_NEON;
4224 GemmMicrokernelTester()
4225 .mr(4)
4226 .nr(8)
4227 .kr(2)
4228 .sr(1)
4229 .m(4)
4230 .n(8)
4231 .k(16)
4232 .cm_stride(11)
4233 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mlal_ld4r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4234 }
4235#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
4236
4237
4238#if XNN_ARCH_ARM || XNN_ARCH_ARM64
4239 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8) {
4240 TEST_REQUIRES_ARM_NEON;
4241 GemmMicrokernelTester()
4242 .mr(1)
4243 .nr(8)
4244 .kr(4)
4245 .sr(2)
4246 .m(1)
4247 .n(8)
4248 .k(8)
4249 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4250 }
4251
4252 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cn) {
4253 TEST_REQUIRES_ARM_NEON;
4254 GemmMicrokernelTester()
4255 .mr(1)
4256 .nr(8)
4257 .kr(4)
4258 .sr(2)
4259 .m(1)
4260 .n(8)
4261 .k(8)
4262 .cn_stride(11)
4263 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4264 }
4265
4266 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile) {
4267 TEST_REQUIRES_ARM_NEON;
4268 for (uint32_t n = 1; n <= 8; n++) {
4269 for (uint32_t m = 1; m <= 1; m++) {
4270 GemmMicrokernelTester()
4271 .mr(1)
4272 .nr(8)
4273 .kr(4)
4274 .sr(2)
4275 .m(m)
4276 .n(n)
4277 .k(8)
4278 .iterations(1)
4279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4280 }
4281 }
4282 }
4283
4284 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile_m) {
4285 TEST_REQUIRES_ARM_NEON;
4286 for (uint32_t m = 1; m <= 1; m++) {
4287 GemmMicrokernelTester()
4288 .mr(1)
4289 .nr(8)
4290 .kr(4)
4291 .sr(2)
4292 .m(m)
4293 .n(8)
4294 .k(8)
4295 .iterations(1)
4296 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4297 }
4298 }
4299
4300 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_eq_8_subtile_n) {
4301 TEST_REQUIRES_ARM_NEON;
4302 for (uint32_t n = 1; n <= 8; n++) {
4303 GemmMicrokernelTester()
4304 .mr(1)
4305 .nr(8)
4306 .kr(4)
4307 .sr(2)
4308 .m(1)
4309 .n(n)
4310 .k(8)
4311 .iterations(1)
4312 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4313 }
4314 }
4315
4316 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_lt_8) {
4317 TEST_REQUIRES_ARM_NEON;
4318 for (size_t k = 1; k < 8; k++) {
4319 GemmMicrokernelTester()
4320 .mr(1)
4321 .nr(8)
4322 .kr(4)
4323 .sr(2)
4324 .m(1)
4325 .n(8)
4326 .k(k)
4327 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4328 }
4329 }
4330
4331 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_lt_8_subtile) {
4332 TEST_REQUIRES_ARM_NEON;
4333 for (size_t k = 1; k < 8; k++) {
4334 for (uint32_t n = 1; n <= 8; n++) {
4335 for (uint32_t m = 1; m <= 1; m++) {
4336 GemmMicrokernelTester()
4337 .mr(1)
4338 .nr(8)
4339 .kr(4)
4340 .sr(2)
4341 .m(m)
4342 .n(n)
4343 .k(k)
4344 .iterations(1)
4345 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4346 }
4347 }
4348 }
4349 }
4350
4351 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_gt_8) {
4352 TEST_REQUIRES_ARM_NEON;
4353 for (size_t k = 9; k < 16; k++) {
4354 GemmMicrokernelTester()
4355 .mr(1)
4356 .nr(8)
4357 .kr(4)
4358 .sr(2)
4359 .m(1)
4360 .n(8)
4361 .k(k)
4362 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4363 }
4364 }
4365
4366 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_gt_8_subtile) {
4367 TEST_REQUIRES_ARM_NEON;
4368 for (size_t k = 9; k < 16; k++) {
4369 for (uint32_t n = 1; n <= 8; n++) {
4370 for (uint32_t m = 1; m <= 1; m++) {
4371 GemmMicrokernelTester()
4372 .mr(1)
4373 .nr(8)
4374 .kr(4)
4375 .sr(2)
4376 .m(m)
4377 .n(n)
4378 .k(k)
4379 .iterations(1)
4380 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4381 }
4382 }
4383 }
4384 }
4385
4386 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_div_8) {
4387 TEST_REQUIRES_ARM_NEON;
4388 for (size_t k = 16; k <= 80; k += 8) {
4389 GemmMicrokernelTester()
4390 .mr(1)
4391 .nr(8)
4392 .kr(4)
4393 .sr(2)
4394 .m(1)
4395 .n(8)
4396 .k(k)
4397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4398 }
4399 }
4400
4401 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, k_div_8_subtile) {
4402 TEST_REQUIRES_ARM_NEON;
4403 for (size_t k = 16; k <= 80; k += 8) {
4404 for (uint32_t n = 1; n <= 8; n++) {
4405 for (uint32_t m = 1; m <= 1; m++) {
4406 GemmMicrokernelTester()
4407 .mr(1)
4408 .nr(8)
4409 .kr(4)
4410 .sr(2)
4411 .m(m)
4412 .n(n)
4413 .k(k)
4414 .iterations(1)
4415 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4416 }
4417 }
4418 }
4419 }
4420
4421 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8) {
4422 TEST_REQUIRES_ARM_NEON;
4423 for (uint32_t n = 9; n < 16; n++) {
4424 for (size_t k = 1; k <= 40; k += 9) {
4425 GemmMicrokernelTester()
4426 .mr(1)
4427 .nr(8)
4428 .kr(4)
4429 .sr(2)
4430 .m(1)
4431 .n(n)
4432 .k(k)
4433 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4434 }
4435 }
4436 }
4437
4438 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8_strided_cn) {
4439 TEST_REQUIRES_ARM_NEON;
4440 for (uint32_t n = 9; n < 16; n++) {
4441 for (size_t k = 1; k <= 40; k += 9) {
4442 GemmMicrokernelTester()
4443 .mr(1)
4444 .nr(8)
4445 .kr(4)
4446 .sr(2)
4447 .m(1)
4448 .n(n)
4449 .k(k)
4450 .cn_stride(11)
4451 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4452 }
4453 }
4454 }
4455
4456 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8_subtile) {
4457 TEST_REQUIRES_ARM_NEON;
4458 for (uint32_t n = 9; n < 16; n++) {
4459 for (size_t k = 1; k <= 40; k += 9) {
4460 for (uint32_t m = 1; m <= 1; m++) {
4461 GemmMicrokernelTester()
4462 .mr(1)
4463 .nr(8)
4464 .kr(4)
4465 .sr(2)
4466 .m(m)
4467 .n(n)
4468 .k(k)
4469 .iterations(1)
4470 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4471 }
4472 }
4473 }
4474 }
4475
4476 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8) {
4477 TEST_REQUIRES_ARM_NEON;
4478 for (uint32_t n = 16; n <= 24; n += 8) {
4479 for (size_t k = 1; k <= 40; k += 9) {
4480 GemmMicrokernelTester()
4481 .mr(1)
4482 .nr(8)
4483 .kr(4)
4484 .sr(2)
4485 .m(1)
4486 .n(n)
4487 .k(k)
4488 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4489 }
4490 }
4491 }
4492
4493 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8_strided_cn) {
4494 TEST_REQUIRES_ARM_NEON;
4495 for (uint32_t n = 16; n <= 24; n += 8) {
4496 for (size_t k = 1; k <= 40; k += 9) {
4497 GemmMicrokernelTester()
4498 .mr(1)
4499 .nr(8)
4500 .kr(4)
4501 .sr(2)
4502 .m(1)
4503 .n(n)
4504 .k(k)
4505 .cn_stride(11)
4506 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4507 }
4508 }
4509 }
4510
4511 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8_subtile) {
4512 TEST_REQUIRES_ARM_NEON;
4513 for (uint32_t n = 16; n <= 24; n += 8) {
4514 for (size_t k = 1; k <= 40; k += 9) {
4515 for (uint32_t m = 1; m <= 1; m++) {
4516 GemmMicrokernelTester()
4517 .mr(1)
4518 .nr(8)
4519 .kr(4)
4520 .sr(2)
4521 .m(m)
4522 .n(n)
4523 .k(k)
4524 .iterations(1)
4525 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4526 }
4527 }
4528 }
4529 }
4530
4531 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, small_kernel) {
4532 TEST_REQUIRES_ARM_NEON;
4533 for (size_t k = 1; k <= 40; k += 9) {
4534 GemmMicrokernelTester()
4535 .mr(1)
4536 .nr(8)
4537 .kr(4)
4538 .sr(2)
4539 .m(1)
4540 .n(8)
4541 .k(k)
4542 .ks(3)
4543 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4544 }
4545 }
4546
4547 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, small_kernel_subtile) {
4548 TEST_REQUIRES_ARM_NEON;
4549 for (size_t k = 1; k <= 40; k += 9) {
4550 for (uint32_t n = 1; n <= 8; n++) {
4551 for (uint32_t m = 1; m <= 1; m++) {
4552 GemmMicrokernelTester()
4553 .mr(1)
4554 .nr(8)
4555 .kr(4)
4556 .sr(2)
4557 .m(m)
4558 .n(n)
4559 .k(k)
4560 .ks(3)
4561 .iterations(1)
4562 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4563 }
4564 }
4565 }
4566 }
4567
4568 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_gt_8_small_kernel) {
4569 TEST_REQUIRES_ARM_NEON;
4570 for (uint32_t n = 9; n < 16; n++) {
4571 for (size_t k = 1; k <= 40; k += 9) {
4572 GemmMicrokernelTester()
4573 .mr(1)
4574 .nr(8)
4575 .kr(4)
4576 .sr(2)
4577 .m(1)
4578 .n(n)
4579 .k(k)
4580 .ks(3)
4581 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4582 }
4583 }
4584 }
4585
4586 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, n_div_8_small_kernel) {
4587 TEST_REQUIRES_ARM_NEON;
4588 for (uint32_t n = 16; n <= 24; n += 8) {
4589 for (size_t k = 1; k <= 40; k += 9) {
4590 GemmMicrokernelTester()
4591 .mr(1)
4592 .nr(8)
4593 .kr(4)
4594 .sr(2)
4595 .m(1)
4596 .n(n)
4597 .k(k)
4598 .ks(3)
4599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4600 }
4601 }
4602 }
4603
4604 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cm_subtile) {
4605 TEST_REQUIRES_ARM_NEON;
4606 for (size_t k = 1; k <= 40; k += 9) {
4607 for (uint32_t n = 1; n <= 8; n++) {
4608 for (uint32_t m = 1; m <= 1; m++) {
4609 GemmMicrokernelTester()
4610 .mr(1)
4611 .nr(8)
4612 .kr(4)
4613 .sr(2)
4614 .m(m)
4615 .n(n)
4616 .k(k)
4617 .cm_stride(11)
4618 .iterations(1)
4619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4620 }
4621 }
4622 }
4623 }
4624
4625 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, a_offset) {
4626 TEST_REQUIRES_ARM_NEON;
4627 for (size_t k = 1; k <= 40; k += 9) {
4628 GemmMicrokernelTester()
4629 .mr(1)
4630 .nr(8)
4631 .kr(4)
4632 .sr(2)
4633 .m(1)
4634 .n(8)
4635 .k(k)
4636 .ks(3)
4637 .a_offset(43)
4638 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4639 }
4640 }
4641
4642 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, zero) {
4643 TEST_REQUIRES_ARM_NEON;
4644 for (size_t k = 1; k <= 40; k += 9) {
4645 for (uint32_t mz = 0; mz < 1; mz++) {
4646 GemmMicrokernelTester()
4647 .mr(1)
4648 .nr(8)
4649 .kr(4)
4650 .sr(2)
4651 .m(1)
4652 .n(8)
4653 .k(k)
4654 .ks(3)
4655 .a_offset(43)
4656 .zero_index(mz)
4657 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4658 }
4659 }
4660 }
4661
4662 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, qmin) {
4663 TEST_REQUIRES_ARM_NEON;
4664 GemmMicrokernelTester()
4665 .mr(1)
4666 .nr(8)
4667 .kr(4)
4668 .sr(2)
4669 .m(1)
4670 .n(8)
4671 .k(8)
4672 .qmin(128)
4673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4674 }
4675
4676 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, qmax) {
4677 TEST_REQUIRES_ARM_NEON;
4678 GemmMicrokernelTester()
4679 .mr(1)
4680 .nr(8)
4681 .kr(4)
4682 .sr(2)
4683 .m(1)
4684 .n(8)
4685 .k(8)
4686 .qmax(128)
4687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4688 }
4689
4690 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4S2__NEON_MULL, strided_cm) {
4691 TEST_REQUIRES_ARM_NEON;
4692 GemmMicrokernelTester()
4693 .mr(1)
4694 .nr(8)
4695 .kr(4)
4696 .sr(2)
4697 .m(1)
4698 .n(8)
4699 .k(8)
4700 .cm_stride(11)
4701 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4702 }
4703#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
4704
4705
4706#if XNN_ARCH_ARM || XNN_ARCH_ARM64
4707 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_eq_8) {
4708 TEST_REQUIRES_ARM_NEON;
4709 GemmMicrokernelTester()
4710 .mr(4)
4711 .nr(16)
4712 .kr(4)
4713 .sr(2)
4714 .m(4)
4715 .n(16)
4716 .k(8)
4717 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4718 }
4719
4720 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, strided_cn) {
4721 TEST_REQUIRES_ARM_NEON;
4722 GemmMicrokernelTester()
4723 .mr(4)
4724 .nr(16)
4725 .kr(4)
4726 .sr(2)
4727 .m(4)
4728 .n(16)
4729 .k(8)
4730 .cn_stride(19)
4731 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4732 }
4733
4734 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_eq_8_subtile) {
4735 TEST_REQUIRES_ARM_NEON;
4736 for (uint32_t n = 1; n <= 16; n++) {
4737 for (uint32_t m = 1; m <= 4; m++) {
4738 GemmMicrokernelTester()
4739 .mr(4)
4740 .nr(16)
4741 .kr(4)
4742 .sr(2)
4743 .m(m)
4744 .n(n)
4745 .k(8)
4746 .iterations(1)
4747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4748 }
4749 }
4750 }
4751
4752 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_eq_8_subtile_m) {
4753 TEST_REQUIRES_ARM_NEON;
4754 for (uint32_t m = 1; m <= 4; m++) {
4755 GemmMicrokernelTester()
4756 .mr(4)
4757 .nr(16)
4758 .kr(4)
4759 .sr(2)
4760 .m(m)
4761 .n(16)
4762 .k(8)
4763 .iterations(1)
4764 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4765 }
4766 }
4767
4768 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_eq_8_subtile_n) {
4769 TEST_REQUIRES_ARM_NEON;
4770 for (uint32_t n = 1; n <= 16; n++) {
4771 GemmMicrokernelTester()
4772 .mr(4)
4773 .nr(16)
4774 .kr(4)
4775 .sr(2)
4776 .m(4)
4777 .n(n)
4778 .k(8)
4779 .iterations(1)
4780 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4781 }
4782 }
4783
4784 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_lt_8) {
4785 TEST_REQUIRES_ARM_NEON;
4786 for (size_t k = 1; k < 8; k++) {
4787 GemmMicrokernelTester()
4788 .mr(4)
4789 .nr(16)
4790 .kr(4)
4791 .sr(2)
4792 .m(4)
4793 .n(16)
4794 .k(k)
4795 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4796 }
4797 }
4798
4799 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_lt_8_subtile) {
4800 TEST_REQUIRES_ARM_NEON;
4801 for (size_t k = 1; k < 8; k++) {
4802 for (uint32_t n = 1; n <= 16; n++) {
4803 for (uint32_t m = 1; m <= 4; m++) {
4804 GemmMicrokernelTester()
4805 .mr(4)
4806 .nr(16)
4807 .kr(4)
4808 .sr(2)
4809 .m(m)
4810 .n(n)
4811 .k(k)
4812 .iterations(1)
4813 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4814 }
4815 }
4816 }
4817 }
4818
4819 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_gt_8) {
4820 TEST_REQUIRES_ARM_NEON;
4821 for (size_t k = 9; k < 16; k++) {
4822 GemmMicrokernelTester()
4823 .mr(4)
4824 .nr(16)
4825 .kr(4)
4826 .sr(2)
4827 .m(4)
4828 .n(16)
4829 .k(k)
4830 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4831 }
4832 }
4833
4834 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_gt_8_subtile) {
4835 TEST_REQUIRES_ARM_NEON;
4836 for (size_t k = 9; k < 16; k++) {
4837 for (uint32_t n = 1; n <= 16; n++) {
4838 for (uint32_t m = 1; m <= 4; m++) {
4839 GemmMicrokernelTester()
4840 .mr(4)
4841 .nr(16)
4842 .kr(4)
4843 .sr(2)
4844 .m(m)
4845 .n(n)
4846 .k(k)
4847 .iterations(1)
4848 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4849 }
4850 }
4851 }
4852 }
4853
4854 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_div_8) {
4855 TEST_REQUIRES_ARM_NEON;
4856 for (size_t k = 16; k <= 80; k += 8) {
4857 GemmMicrokernelTester()
4858 .mr(4)
4859 .nr(16)
4860 .kr(4)
4861 .sr(2)
4862 .m(4)
4863 .n(16)
4864 .k(k)
4865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4866 }
4867 }
4868
4869 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, k_div_8_subtile) {
4870 TEST_REQUIRES_ARM_NEON;
4871 for (size_t k = 16; k <= 80; k += 8) {
4872 for (uint32_t n = 1; n <= 16; n++) {
4873 for (uint32_t m = 1; m <= 4; m++) {
4874 GemmMicrokernelTester()
4875 .mr(4)
4876 .nr(16)
4877 .kr(4)
4878 .sr(2)
4879 .m(m)
4880 .n(n)
4881 .k(k)
4882 .iterations(1)
4883 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4884 }
4885 }
4886 }
4887 }
4888
4889 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_gt_16) {
4890 TEST_REQUIRES_ARM_NEON;
4891 for (uint32_t n = 17; n < 32; n++) {
4892 for (size_t k = 1; k <= 40; k += 9) {
4893 GemmMicrokernelTester()
4894 .mr(4)
4895 .nr(16)
4896 .kr(4)
4897 .sr(2)
4898 .m(4)
4899 .n(n)
4900 .k(k)
4901 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4902 }
4903 }
4904 }
4905
4906 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_gt_16_strided_cn) {
4907 TEST_REQUIRES_ARM_NEON;
4908 for (uint32_t n = 17; n < 32; n++) {
4909 for (size_t k = 1; k <= 40; k += 9) {
4910 GemmMicrokernelTester()
4911 .mr(4)
4912 .nr(16)
4913 .kr(4)
4914 .sr(2)
4915 .m(4)
4916 .n(n)
4917 .k(k)
4918 .cn_stride(19)
4919 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4920 }
4921 }
4922 }
4923
4924 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_gt_16_subtile) {
4925 TEST_REQUIRES_ARM_NEON;
4926 for (uint32_t n = 17; n < 32; n++) {
4927 for (size_t k = 1; k <= 40; k += 9) {
4928 for (uint32_t m = 1; m <= 4; m++) {
4929 GemmMicrokernelTester()
4930 .mr(4)
4931 .nr(16)
4932 .kr(4)
4933 .sr(2)
4934 .m(m)
4935 .n(n)
4936 .k(k)
4937 .iterations(1)
4938 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4939 }
4940 }
4941 }
4942 }
4943
4944 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_div_16) {
4945 TEST_REQUIRES_ARM_NEON;
4946 for (uint32_t n = 32; n <= 48; n += 16) {
4947 for (size_t k = 1; k <= 40; k += 9) {
4948 GemmMicrokernelTester()
4949 .mr(4)
4950 .nr(16)
4951 .kr(4)
4952 .sr(2)
4953 .m(4)
4954 .n(n)
4955 .k(k)
4956 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4957 }
4958 }
4959 }
4960
4961 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_div_16_strided_cn) {
4962 TEST_REQUIRES_ARM_NEON;
4963 for (uint32_t n = 32; n <= 48; n += 16) {
4964 for (size_t k = 1; k <= 40; k += 9) {
4965 GemmMicrokernelTester()
4966 .mr(4)
4967 .nr(16)
4968 .kr(4)
4969 .sr(2)
4970 .m(4)
4971 .n(n)
4972 .k(k)
4973 .cn_stride(19)
4974 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4975 }
4976 }
4977 }
4978
4979 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_div_16_subtile) {
4980 TEST_REQUIRES_ARM_NEON;
4981 for (uint32_t n = 32; n <= 48; n += 16) {
4982 for (size_t k = 1; k <= 40; k += 9) {
4983 for (uint32_t m = 1; m <= 4; m++) {
4984 GemmMicrokernelTester()
4985 .mr(4)
4986 .nr(16)
4987 .kr(4)
4988 .sr(2)
4989 .m(m)
4990 .n(n)
4991 .k(k)
4992 .iterations(1)
4993 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
4994 }
4995 }
4996 }
4997 }
4998
4999 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, small_kernel) {
5000 TEST_REQUIRES_ARM_NEON;
5001 for (size_t k = 1; k <= 40; k += 9) {
5002 GemmMicrokernelTester()
5003 .mr(4)
5004 .nr(16)
5005 .kr(4)
5006 .sr(2)
5007 .m(4)
5008 .n(16)
5009 .k(k)
5010 .ks(3)
5011 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5012 }
5013 }
5014
5015 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, small_kernel_subtile) {
5016 TEST_REQUIRES_ARM_NEON;
5017 for (size_t k = 1; k <= 40; k += 9) {
5018 for (uint32_t n = 1; n <= 16; n++) {
5019 for (uint32_t m = 1; m <= 4; m++) {
5020 GemmMicrokernelTester()
5021 .mr(4)
5022 .nr(16)
5023 .kr(4)
5024 .sr(2)
5025 .m(m)
5026 .n(n)
5027 .k(k)
5028 .ks(3)
5029 .iterations(1)
5030 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5031 }
5032 }
5033 }
5034 }
5035
5036 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_gt_16_small_kernel) {
5037 TEST_REQUIRES_ARM_NEON;
5038 for (uint32_t n = 17; n < 32; n++) {
5039 for (size_t k = 1; k <= 40; k += 9) {
5040 GemmMicrokernelTester()
5041 .mr(4)
5042 .nr(16)
5043 .kr(4)
5044 .sr(2)
5045 .m(4)
5046 .n(n)
5047 .k(k)
5048 .ks(3)
5049 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5050 }
5051 }
5052 }
5053
5054 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, n_div_16_small_kernel) {
5055 TEST_REQUIRES_ARM_NEON;
5056 for (uint32_t n = 32; n <= 48; n += 16) {
5057 for (size_t k = 1; k <= 40; k += 9) {
5058 GemmMicrokernelTester()
5059 .mr(4)
5060 .nr(16)
5061 .kr(4)
5062 .sr(2)
5063 .m(4)
5064 .n(n)
5065 .k(k)
5066 .ks(3)
5067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5068 }
5069 }
5070 }
5071
5072 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, strided_cm_subtile) {
5073 TEST_REQUIRES_ARM_NEON;
5074 for (size_t k = 1; k <= 40; k += 9) {
5075 for (uint32_t n = 1; n <= 16; n++) {
5076 for (uint32_t m = 1; m <= 4; m++) {
5077 GemmMicrokernelTester()
5078 .mr(4)
5079 .nr(16)
5080 .kr(4)
5081 .sr(2)
5082 .m(m)
5083 .n(n)
5084 .k(k)
5085 .cm_stride(19)
5086 .iterations(1)
5087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5088 }
5089 }
5090 }
5091 }
5092
5093 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, a_offset) {
5094 TEST_REQUIRES_ARM_NEON;
5095 for (size_t k = 1; k <= 40; k += 9) {
5096 GemmMicrokernelTester()
5097 .mr(4)
5098 .nr(16)
5099 .kr(4)
5100 .sr(2)
5101 .m(4)
5102 .n(16)
5103 .k(k)
5104 .ks(3)
5105 .a_offset(163)
5106 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5107 }
5108 }
5109
5110 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, zero) {
5111 TEST_REQUIRES_ARM_NEON;
5112 for (size_t k = 1; k <= 40; k += 9) {
5113 for (uint32_t mz = 0; mz < 4; mz++) {
5114 GemmMicrokernelTester()
5115 .mr(4)
5116 .nr(16)
5117 .kr(4)
5118 .sr(2)
5119 .m(4)
5120 .n(16)
5121 .k(k)
5122 .ks(3)
5123 .a_offset(163)
5124 .zero_index(mz)
5125 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5126 }
5127 }
5128 }
5129
5130 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, qmin) {
5131 TEST_REQUIRES_ARM_NEON;
5132 GemmMicrokernelTester()
5133 .mr(4)
5134 .nr(16)
5135 .kr(4)
5136 .sr(2)
5137 .m(4)
5138 .n(16)
5139 .k(8)
5140 .qmin(128)
5141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5142 }
5143
5144 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, qmax) {
5145 TEST_REQUIRES_ARM_NEON;
5146 GemmMicrokernelTester()
5147 .mr(4)
5148 .nr(16)
5149 .kr(4)
5150 .sr(2)
5151 .m(4)
5152 .n(16)
5153 .k(8)
5154 .qmax(128)
5155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5156 }
5157
5158 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4S2__NEON_MULL, strided_cm) {
5159 TEST_REQUIRES_ARM_NEON;
5160 GemmMicrokernelTester()
5161 .mr(4)
5162 .nr(16)
5163 .kr(4)
5164 .sr(2)
5165 .m(4)
5166 .n(16)
5167 .k(8)
5168 .cm_stride(19)
5169 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4s2__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5170 }
5171#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
5172
5173
5174#if XNN_ARCH_ARM || XNN_ARCH_ARM64
5175 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_eq_16) {
5176 TEST_REQUIRES_ARM_NEON;
5177 GemmMicrokernelTester()
5178 .mr(2)
5179 .nr(8)
5180 .kr(4)
5181 .sr(2)
5182 .m(2)
5183 .n(8)
5184 .k(16)
5185 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5186 }
5187
5188 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, strided_cn) {
5189 TEST_REQUIRES_ARM_NEON;
5190 GemmMicrokernelTester()
5191 .mr(2)
5192 .nr(8)
5193 .kr(4)
5194 .sr(2)
5195 .m(2)
5196 .n(8)
5197 .k(16)
5198 .cn_stride(11)
5199 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5200 }
5201
5202 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_eq_16_subtile) {
5203 TEST_REQUIRES_ARM_NEON;
5204 for (uint32_t n = 1; n <= 8; n++) {
5205 for (uint32_t m = 1; m <= 2; m++) {
5206 GemmMicrokernelTester()
5207 .mr(2)
5208 .nr(8)
5209 .kr(4)
5210 .sr(2)
5211 .m(m)
5212 .n(n)
5213 .k(16)
5214 .iterations(1)
5215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5216 }
5217 }
5218 }
5219
5220 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_eq_16_subtile_m) {
5221 TEST_REQUIRES_ARM_NEON;
5222 for (uint32_t m = 1; m <= 2; m++) {
5223 GemmMicrokernelTester()
5224 .mr(2)
5225 .nr(8)
5226 .kr(4)
5227 .sr(2)
5228 .m(m)
5229 .n(8)
5230 .k(16)
5231 .iterations(1)
5232 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5233 }
5234 }
5235
5236 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_eq_16_subtile_n) {
5237 TEST_REQUIRES_ARM_NEON;
5238 for (uint32_t n = 1; n <= 8; n++) {
5239 GemmMicrokernelTester()
5240 .mr(2)
5241 .nr(8)
5242 .kr(4)
5243 .sr(2)
5244 .m(2)
5245 .n(n)
5246 .k(16)
5247 .iterations(1)
5248 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5249 }
5250 }
5251
5252 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_lt_16) {
5253 TEST_REQUIRES_ARM_NEON;
5254 for (size_t k = 1; k < 16; k++) {
5255 GemmMicrokernelTester()
5256 .mr(2)
5257 .nr(8)
5258 .kr(4)
5259 .sr(2)
5260 .m(2)
5261 .n(8)
5262 .k(k)
5263 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5264 }
5265 }
5266
5267 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_lt_16_subtile) {
5268 TEST_REQUIRES_ARM_NEON;
5269 for (size_t k = 1; k < 16; k++) {
5270 for (uint32_t n = 1; n <= 8; n++) {
5271 for (uint32_t m = 1; m <= 2; m++) {
5272 GemmMicrokernelTester()
5273 .mr(2)
5274 .nr(8)
5275 .kr(4)
5276 .sr(2)
5277 .m(m)
5278 .n(n)
5279 .k(k)
5280 .iterations(1)
5281 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5282 }
5283 }
5284 }
5285 }
5286
5287 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_gt_16) {
5288 TEST_REQUIRES_ARM_NEON;
5289 for (size_t k = 17; k < 32; k++) {
5290 GemmMicrokernelTester()
5291 .mr(2)
5292 .nr(8)
5293 .kr(4)
5294 .sr(2)
5295 .m(2)
5296 .n(8)
5297 .k(k)
5298 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5299 }
5300 }
5301
5302 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_gt_16_subtile) {
5303 TEST_REQUIRES_ARM_NEON;
5304 for (size_t k = 17; k < 32; k++) {
5305 for (uint32_t n = 1; n <= 8; n++) {
5306 for (uint32_t m = 1; m <= 2; m++) {
5307 GemmMicrokernelTester()
5308 .mr(2)
5309 .nr(8)
5310 .kr(4)
5311 .sr(2)
5312 .m(m)
5313 .n(n)
5314 .k(k)
5315 .iterations(1)
5316 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5317 }
5318 }
5319 }
5320 }
5321
5322 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_div_16) {
5323 TEST_REQUIRES_ARM_NEON;
5324 for (size_t k = 32; k <= 160; k += 16) {
5325 GemmMicrokernelTester()
5326 .mr(2)
5327 .nr(8)
5328 .kr(4)
5329 .sr(2)
5330 .m(2)
5331 .n(8)
5332 .k(k)
5333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5334 }
5335 }
5336
5337 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, k_div_16_subtile) {
5338 TEST_REQUIRES_ARM_NEON;
5339 for (size_t k = 32; k <= 160; k += 16) {
5340 for (uint32_t n = 1; n <= 8; n++) {
5341 for (uint32_t m = 1; m <= 2; m++) {
5342 GemmMicrokernelTester()
5343 .mr(2)
5344 .nr(8)
5345 .kr(4)
5346 .sr(2)
5347 .m(m)
5348 .n(n)
5349 .k(k)
5350 .iterations(1)
5351 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5352 }
5353 }
5354 }
5355 }
5356
5357 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_gt_8) {
5358 TEST_REQUIRES_ARM_NEON;
5359 for (uint32_t n = 9; n < 16; n++) {
5360 for (size_t k = 1; k <= 80; k += 17) {
5361 GemmMicrokernelTester()
5362 .mr(2)
5363 .nr(8)
5364 .kr(4)
5365 .sr(2)
5366 .m(2)
5367 .n(n)
5368 .k(k)
5369 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5370 }
5371 }
5372 }
5373
5374 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_gt_8_strided_cn) {
5375 TEST_REQUIRES_ARM_NEON;
5376 for (uint32_t n = 9; n < 16; n++) {
5377 for (size_t k = 1; k <= 80; k += 17) {
5378 GemmMicrokernelTester()
5379 .mr(2)
5380 .nr(8)
5381 .kr(4)
5382 .sr(2)
5383 .m(2)
5384 .n(n)
5385 .k(k)
5386 .cn_stride(11)
5387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5388 }
5389 }
5390 }
5391
5392 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_gt_8_subtile) {
5393 TEST_REQUIRES_ARM_NEON;
5394 for (uint32_t n = 9; n < 16; n++) {
5395 for (size_t k = 1; k <= 80; k += 17) {
5396 for (uint32_t m = 1; m <= 2; m++) {
5397 GemmMicrokernelTester()
5398 .mr(2)
5399 .nr(8)
5400 .kr(4)
5401 .sr(2)
5402 .m(m)
5403 .n(n)
5404 .k(k)
5405 .iterations(1)
5406 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5407 }
5408 }
5409 }
5410 }
5411
5412 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_div_8) {
5413 TEST_REQUIRES_ARM_NEON;
5414 for (uint32_t n = 16; n <= 24; n += 8) {
5415 for (size_t k = 1; k <= 80; k += 17) {
5416 GemmMicrokernelTester()
5417 .mr(2)
5418 .nr(8)
5419 .kr(4)
5420 .sr(2)
5421 .m(2)
5422 .n(n)
5423 .k(k)
5424 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5425 }
5426 }
5427 }
5428
5429 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_div_8_strided_cn) {
5430 TEST_REQUIRES_ARM_NEON;
5431 for (uint32_t n = 16; n <= 24; n += 8) {
5432 for (size_t k = 1; k <= 80; k += 17) {
5433 GemmMicrokernelTester()
5434 .mr(2)
5435 .nr(8)
5436 .kr(4)
5437 .sr(2)
5438 .m(2)
5439 .n(n)
5440 .k(k)
5441 .cn_stride(11)
5442 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5443 }
5444 }
5445 }
5446
5447 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_div_8_subtile) {
5448 TEST_REQUIRES_ARM_NEON;
5449 for (uint32_t n = 16; n <= 24; n += 8) {
5450 for (size_t k = 1; k <= 80; k += 17) {
5451 for (uint32_t m = 1; m <= 2; m++) {
5452 GemmMicrokernelTester()
5453 .mr(2)
5454 .nr(8)
5455 .kr(4)
5456 .sr(2)
5457 .m(m)
5458 .n(n)
5459 .k(k)
5460 .iterations(1)
5461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5462 }
5463 }
5464 }
5465 }
5466
5467 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, small_kernel) {
5468 TEST_REQUIRES_ARM_NEON;
5469 for (size_t k = 1; k <= 80; k += 17) {
5470 GemmMicrokernelTester()
5471 .mr(2)
5472 .nr(8)
5473 .kr(4)
5474 .sr(2)
5475 .m(2)
5476 .n(8)
5477 .k(k)
5478 .ks(3)
5479 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5480 }
5481 }
5482
5483 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, small_kernel_subtile) {
5484 TEST_REQUIRES_ARM_NEON;
5485 for (size_t k = 1; k <= 80; k += 17) {
5486 for (uint32_t n = 1; n <= 8; n++) {
5487 for (uint32_t m = 1; m <= 2; m++) {
5488 GemmMicrokernelTester()
5489 .mr(2)
5490 .nr(8)
5491 .kr(4)
5492 .sr(2)
5493 .m(m)
5494 .n(n)
5495 .k(k)
5496 .ks(3)
5497 .iterations(1)
5498 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5499 }
5500 }
5501 }
5502 }
5503
5504 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_gt_8_small_kernel) {
5505 TEST_REQUIRES_ARM_NEON;
5506 for (uint32_t n = 9; n < 16; n++) {
5507 for (size_t k = 1; k <= 80; k += 17) {
5508 GemmMicrokernelTester()
5509 .mr(2)
5510 .nr(8)
5511 .kr(4)
5512 .sr(2)
5513 .m(2)
5514 .n(n)
5515 .k(k)
5516 .ks(3)
5517 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5518 }
5519 }
5520 }
5521
5522 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, n_div_8_small_kernel) {
5523 TEST_REQUIRES_ARM_NEON;
5524 for (uint32_t n = 16; n <= 24; n += 8) {
5525 for (size_t k = 1; k <= 80; k += 17) {
5526 GemmMicrokernelTester()
5527 .mr(2)
5528 .nr(8)
5529 .kr(4)
5530 .sr(2)
5531 .m(2)
5532 .n(n)
5533 .k(k)
5534 .ks(3)
5535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5536 }
5537 }
5538 }
5539
5540 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, strided_cm_subtile) {
5541 TEST_REQUIRES_ARM_NEON;
5542 for (size_t k = 1; k <= 80; k += 17) {
5543 for (uint32_t n = 1; n <= 8; n++) {
5544 for (uint32_t m = 1; m <= 2; m++) {
5545 GemmMicrokernelTester()
5546 .mr(2)
5547 .nr(8)
5548 .kr(4)
5549 .sr(2)
5550 .m(m)
5551 .n(n)
5552 .k(k)
5553 .cm_stride(11)
5554 .iterations(1)
5555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5556 }
5557 }
5558 }
5559 }
5560
5561 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, a_offset) {
5562 TEST_REQUIRES_ARM_NEON;
5563 for (size_t k = 1; k <= 80; k += 17) {
5564 GemmMicrokernelTester()
5565 .mr(2)
5566 .nr(8)
5567 .kr(4)
5568 .sr(2)
5569 .m(2)
5570 .n(8)
5571 .k(k)
5572 .ks(3)
5573 .a_offset(163)
5574 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5575 }
5576 }
5577
5578 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, zero) {
5579 TEST_REQUIRES_ARM_NEON;
5580 for (size_t k = 1; k <= 80; k += 17) {
5581 for (uint32_t mz = 0; mz < 2; mz++) {
5582 GemmMicrokernelTester()
5583 .mr(2)
5584 .nr(8)
5585 .kr(4)
5586 .sr(2)
5587 .m(2)
5588 .n(8)
5589 .k(k)
5590 .ks(3)
5591 .a_offset(163)
5592 .zero_index(mz)
5593 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5594 }
5595 }
5596 }
5597
5598 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, qmin) {
5599 TEST_REQUIRES_ARM_NEON;
5600 GemmMicrokernelTester()
5601 .mr(2)
5602 .nr(8)
5603 .kr(4)
5604 .sr(2)
5605 .m(2)
5606 .n(8)
5607 .k(16)
5608 .qmin(128)
5609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5610 }
5611
5612 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, qmax) {
5613 TEST_REQUIRES_ARM_NEON;
5614 GemmMicrokernelTester()
5615 .mr(2)
5616 .nr(8)
5617 .kr(4)
5618 .sr(2)
5619 .m(2)
5620 .n(8)
5621 .k(16)
5622 .qmax(128)
5623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5624 }
5625
5626 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4S2__NEON_MLAL, strided_cm) {
5627 TEST_REQUIRES_ARM_NEON;
5628 GemmMicrokernelTester()
5629 .mr(2)
5630 .nr(8)
5631 .kr(4)
5632 .sr(2)
5633 .m(2)
5634 .n(8)
5635 .k(16)
5636 .cm_stride(11)
5637 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5638 }
5639#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
5640
5641
5642#if XNN_ARCH_ARM || XNN_ARCH_ARM64
5643 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16) {
5644 TEST_REQUIRES_ARM_NEON;
5645 GemmMicrokernelTester()
5646 .mr(1)
5647 .nr(16)
5648 .kr(4)
5649 .sr(2)
5650 .m(1)
5651 .n(16)
5652 .k(16)
5653 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5654 }
5655
5656 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, strided_cn) {
5657 TEST_REQUIRES_ARM_NEON;
5658 GemmMicrokernelTester()
5659 .mr(1)
5660 .nr(16)
5661 .kr(4)
5662 .sr(2)
5663 .m(1)
5664 .n(16)
5665 .k(16)
5666 .cn_stride(19)
5667 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5668 }
5669
5670 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16_subtile) {
5671 TEST_REQUIRES_ARM_NEON;
5672 for (uint32_t n = 1; n <= 16; n++) {
5673 for (uint32_t m = 1; m <= 1; m++) {
5674 GemmMicrokernelTester()
5675 .mr(1)
5676 .nr(16)
5677 .kr(4)
5678 .sr(2)
5679 .m(m)
5680 .n(n)
5681 .k(16)
5682 .iterations(1)
5683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5684 }
5685 }
5686 }
5687
5688 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16_subtile_m) {
5689 TEST_REQUIRES_ARM_NEON;
5690 for (uint32_t m = 1; m <= 1; m++) {
5691 GemmMicrokernelTester()
5692 .mr(1)
5693 .nr(16)
5694 .kr(4)
5695 .sr(2)
5696 .m(m)
5697 .n(16)
5698 .k(16)
5699 .iterations(1)
5700 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5701 }
5702 }
5703
5704 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_eq_16_subtile_n) {
5705 TEST_REQUIRES_ARM_NEON;
5706 for (uint32_t n = 1; n <= 16; n++) {
5707 GemmMicrokernelTester()
5708 .mr(1)
5709 .nr(16)
5710 .kr(4)
5711 .sr(2)
5712 .m(1)
5713 .n(n)
5714 .k(16)
5715 .iterations(1)
5716 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5717 }
5718 }
5719
5720 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_lt_16) {
5721 TEST_REQUIRES_ARM_NEON;
5722 for (size_t k = 1; k < 16; k++) {
5723 GemmMicrokernelTester()
5724 .mr(1)
5725 .nr(16)
5726 .kr(4)
5727 .sr(2)
5728 .m(1)
5729 .n(16)
5730 .k(k)
5731 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5732 }
5733 }
5734
5735 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_lt_16_subtile) {
5736 TEST_REQUIRES_ARM_NEON;
5737 for (size_t k = 1; k < 16; k++) {
5738 for (uint32_t n = 1; n <= 16; n++) {
5739 for (uint32_t m = 1; m <= 1; m++) {
5740 GemmMicrokernelTester()
5741 .mr(1)
5742 .nr(16)
5743 .kr(4)
5744 .sr(2)
5745 .m(m)
5746 .n(n)
5747 .k(k)
5748 .iterations(1)
5749 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5750 }
5751 }
5752 }
5753 }
5754
5755 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_gt_16) {
5756 TEST_REQUIRES_ARM_NEON;
5757 for (size_t k = 17; k < 32; k++) {
5758 GemmMicrokernelTester()
5759 .mr(1)
5760 .nr(16)
5761 .kr(4)
5762 .sr(2)
5763 .m(1)
5764 .n(16)
5765 .k(k)
5766 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5767 }
5768 }
5769
5770 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_gt_16_subtile) {
5771 TEST_REQUIRES_ARM_NEON;
5772 for (size_t k = 17; k < 32; k++) {
5773 for (uint32_t n = 1; n <= 16; n++) {
5774 for (uint32_t m = 1; m <= 1; m++) {
5775 GemmMicrokernelTester()
5776 .mr(1)
5777 .nr(16)
5778 .kr(4)
5779 .sr(2)
5780 .m(m)
5781 .n(n)
5782 .k(k)
5783 .iterations(1)
5784 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5785 }
5786 }
5787 }
5788 }
5789
5790 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_div_16) {
5791 TEST_REQUIRES_ARM_NEON;
5792 for (size_t k = 32; k <= 160; k += 16) {
5793 GemmMicrokernelTester()
5794 .mr(1)
5795 .nr(16)
5796 .kr(4)
5797 .sr(2)
5798 .m(1)
5799 .n(16)
5800 .k(k)
5801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5802 }
5803 }
5804
5805 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, k_div_16_subtile) {
5806 TEST_REQUIRES_ARM_NEON;
5807 for (size_t k = 32; k <= 160; k += 16) {
5808 for (uint32_t n = 1; n <= 16; n++) {
5809 for (uint32_t m = 1; m <= 1; m++) {
5810 GemmMicrokernelTester()
5811 .mr(1)
5812 .nr(16)
5813 .kr(4)
5814 .sr(2)
5815 .m(m)
5816 .n(n)
5817 .k(k)
5818 .iterations(1)
5819 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5820 }
5821 }
5822 }
5823 }
5824
5825 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16) {
5826 TEST_REQUIRES_ARM_NEON;
5827 for (uint32_t n = 17; n < 32; n++) {
5828 for (size_t k = 1; k <= 80; k += 17) {
5829 GemmMicrokernelTester()
5830 .mr(1)
5831 .nr(16)
5832 .kr(4)
5833 .sr(2)
5834 .m(1)
5835 .n(n)
5836 .k(k)
5837 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5838 }
5839 }
5840 }
5841
5842 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16_strided_cn) {
5843 TEST_REQUIRES_ARM_NEON;
5844 for (uint32_t n = 17; n < 32; n++) {
5845 for (size_t k = 1; k <= 80; k += 17) {
5846 GemmMicrokernelTester()
5847 .mr(1)
5848 .nr(16)
5849 .kr(4)
5850 .sr(2)
5851 .m(1)
5852 .n(n)
5853 .k(k)
5854 .cn_stride(19)
5855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5856 }
5857 }
5858 }
5859
5860 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16_subtile) {
5861 TEST_REQUIRES_ARM_NEON;
5862 for (uint32_t n = 17; n < 32; n++) {
5863 for (size_t k = 1; k <= 80; k += 17) {
5864 for (uint32_t m = 1; m <= 1; m++) {
5865 GemmMicrokernelTester()
5866 .mr(1)
5867 .nr(16)
5868 .kr(4)
5869 .sr(2)
5870 .m(m)
5871 .n(n)
5872 .k(k)
5873 .iterations(1)
5874 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5875 }
5876 }
5877 }
5878 }
5879
5880 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16) {
5881 TEST_REQUIRES_ARM_NEON;
5882 for (uint32_t n = 32; n <= 48; n += 16) {
5883 for (size_t k = 1; k <= 80; k += 17) {
5884 GemmMicrokernelTester()
5885 .mr(1)
5886 .nr(16)
5887 .kr(4)
5888 .sr(2)
5889 .m(1)
5890 .n(n)
5891 .k(k)
5892 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5893 }
5894 }
5895 }
5896
5897 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16_strided_cn) {
5898 TEST_REQUIRES_ARM_NEON;
5899 for (uint32_t n = 32; n <= 48; n += 16) {
5900 for (size_t k = 1; k <= 80; k += 17) {
5901 GemmMicrokernelTester()
5902 .mr(1)
5903 .nr(16)
5904 .kr(4)
5905 .sr(2)
5906 .m(1)
5907 .n(n)
5908 .k(k)
5909 .cn_stride(19)
5910 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5911 }
5912 }
5913 }
5914
5915 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16_subtile) {
5916 TEST_REQUIRES_ARM_NEON;
5917 for (uint32_t n = 32; n <= 48; n += 16) {
5918 for (size_t k = 1; k <= 80; k += 17) {
5919 for (uint32_t m = 1; m <= 1; m++) {
5920 GemmMicrokernelTester()
5921 .mr(1)
5922 .nr(16)
5923 .kr(4)
5924 .sr(2)
5925 .m(m)
5926 .n(n)
5927 .k(k)
5928 .iterations(1)
5929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5930 }
5931 }
5932 }
5933 }
5934
5935 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, small_kernel) {
5936 TEST_REQUIRES_ARM_NEON;
5937 for (size_t k = 1; k <= 80; k += 17) {
5938 GemmMicrokernelTester()
5939 .mr(1)
5940 .nr(16)
5941 .kr(4)
5942 .sr(2)
5943 .m(1)
5944 .n(16)
5945 .k(k)
5946 .ks(3)
5947 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5948 }
5949 }
5950
5951 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, small_kernel_subtile) {
5952 TEST_REQUIRES_ARM_NEON;
5953 for (size_t k = 1; k <= 80; k += 17) {
5954 for (uint32_t n = 1; n <= 16; n++) {
5955 for (uint32_t m = 1; m <= 1; m++) {
5956 GemmMicrokernelTester()
5957 .mr(1)
5958 .nr(16)
5959 .kr(4)
5960 .sr(2)
5961 .m(m)
5962 .n(n)
5963 .k(k)
5964 .ks(3)
5965 .iterations(1)
5966 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5967 }
5968 }
5969 }
5970 }
5971
5972 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_gt_16_small_kernel) {
5973 TEST_REQUIRES_ARM_NEON;
5974 for (uint32_t n = 17; n < 32; n++) {
5975 for (size_t k = 1; k <= 80; k += 17) {
5976 GemmMicrokernelTester()
5977 .mr(1)
5978 .nr(16)
5979 .kr(4)
5980 .sr(2)
5981 .m(1)
5982 .n(n)
5983 .k(k)
5984 .ks(3)
5985 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
5986 }
5987 }
5988 }
5989
5990 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, n_div_16_small_kernel) {
5991 TEST_REQUIRES_ARM_NEON;
5992 for (uint32_t n = 32; n <= 48; n += 16) {
5993 for (size_t k = 1; k <= 80; k += 17) {
5994 GemmMicrokernelTester()
5995 .mr(1)
5996 .nr(16)
5997 .kr(4)
5998 .sr(2)
5999 .m(1)
6000 .n(n)
6001 .k(k)
6002 .ks(3)
6003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6004 }
6005 }
6006 }
6007
6008 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, strided_cm_subtile) {
6009 TEST_REQUIRES_ARM_NEON;
6010 for (size_t k = 1; k <= 80; k += 17) {
6011 for (uint32_t n = 1; n <= 16; n++) {
6012 for (uint32_t m = 1; m <= 1; m++) {
6013 GemmMicrokernelTester()
6014 .mr(1)
6015 .nr(16)
6016 .kr(4)
6017 .sr(2)
6018 .m(m)
6019 .n(n)
6020 .k(k)
6021 .cm_stride(19)
6022 .iterations(1)
6023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6024 }
6025 }
6026 }
6027 }
6028
6029 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, a_offset) {
6030 TEST_REQUIRES_ARM_NEON;
6031 for (size_t k = 1; k <= 80; k += 17) {
6032 GemmMicrokernelTester()
6033 .mr(1)
6034 .nr(16)
6035 .kr(4)
6036 .sr(2)
6037 .m(1)
6038 .n(16)
6039 .k(k)
6040 .ks(3)
6041 .a_offset(83)
6042 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6043 }
6044 }
6045
6046 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, zero) {
6047 TEST_REQUIRES_ARM_NEON;
6048 for (size_t k = 1; k <= 80; k += 17) {
6049 for (uint32_t mz = 0; mz < 1; mz++) {
6050 GemmMicrokernelTester()
6051 .mr(1)
6052 .nr(16)
6053 .kr(4)
6054 .sr(2)
6055 .m(1)
6056 .n(16)
6057 .k(k)
6058 .ks(3)
6059 .a_offset(83)
6060 .zero_index(mz)
6061 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6062 }
6063 }
6064 }
6065
6066 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, qmin) {
6067 TEST_REQUIRES_ARM_NEON;
6068 GemmMicrokernelTester()
6069 .mr(1)
6070 .nr(16)
6071 .kr(4)
6072 .sr(2)
6073 .m(1)
6074 .n(16)
6075 .k(16)
6076 .qmin(128)
6077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6078 }
6079
6080 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, qmax) {
6081 TEST_REQUIRES_ARM_NEON;
6082 GemmMicrokernelTester()
6083 .mr(1)
6084 .nr(16)
6085 .kr(4)
6086 .sr(2)
6087 .m(1)
6088 .n(16)
6089 .k(16)
6090 .qmax(128)
6091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6092 }
6093
6094 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4S2__NEON_MLAL, strided_cm) {
6095 TEST_REQUIRES_ARM_NEON;
6096 GemmMicrokernelTester()
6097 .mr(1)
6098 .nr(16)
6099 .kr(4)
6100 .sr(2)
6101 .m(1)
6102 .n(16)
6103 .k(16)
6104 .cm_stride(19)
6105 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4s2__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6106 }
6107#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
6108
6109
6110#if XNN_ARCH_ARM || XNN_ARCH_ARM64
6111 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_eq_8) {
6112 TEST_REQUIRES_ARM_NEON;
6113 GemmMicrokernelTester()
6114 .mr(2)
6115 .nr(8)
6116 .kr(2)
6117 .sr(4)
6118 .m(2)
6119 .n(8)
6120 .k(8)
6121 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6122 }
6123
6124 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, strided_cn) {
6125 TEST_REQUIRES_ARM_NEON;
6126 GemmMicrokernelTester()
6127 .mr(2)
6128 .nr(8)
6129 .kr(2)
6130 .sr(4)
6131 .m(2)
6132 .n(8)
6133 .k(8)
6134 .cn_stride(11)
6135 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6136 }
6137
6138 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_eq_8_subtile) {
6139 TEST_REQUIRES_ARM_NEON;
6140 for (uint32_t n = 1; n <= 8; n++) {
6141 for (uint32_t m = 1; m <= 2; m++) {
6142 GemmMicrokernelTester()
6143 .mr(2)
6144 .nr(8)
6145 .kr(2)
6146 .sr(4)
6147 .m(m)
6148 .n(n)
6149 .k(8)
6150 .iterations(1)
6151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6152 }
6153 }
6154 }
6155
6156 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_eq_8_subtile_m) {
6157 TEST_REQUIRES_ARM_NEON;
6158 for (uint32_t m = 1; m <= 2; m++) {
6159 GemmMicrokernelTester()
6160 .mr(2)
6161 .nr(8)
6162 .kr(2)
6163 .sr(4)
6164 .m(m)
6165 .n(8)
6166 .k(8)
6167 .iterations(1)
6168 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6169 }
6170 }
6171
6172 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_eq_8_subtile_n) {
6173 TEST_REQUIRES_ARM_NEON;
6174 for (uint32_t n = 1; n <= 8; n++) {
6175 GemmMicrokernelTester()
6176 .mr(2)
6177 .nr(8)
6178 .kr(2)
6179 .sr(4)
6180 .m(2)
6181 .n(n)
6182 .k(8)
6183 .iterations(1)
6184 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6185 }
6186 }
6187
6188 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_lt_8) {
6189 TEST_REQUIRES_ARM_NEON;
6190 for (size_t k = 1; k < 8; k++) {
6191 GemmMicrokernelTester()
6192 .mr(2)
6193 .nr(8)
6194 .kr(2)
6195 .sr(4)
6196 .m(2)
6197 .n(8)
6198 .k(k)
6199 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6200 }
6201 }
6202
6203 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_lt_8_subtile) {
6204 TEST_REQUIRES_ARM_NEON;
6205 for (size_t k = 1; k < 8; k++) {
6206 for (uint32_t n = 1; n <= 8; n++) {
6207 for (uint32_t m = 1; m <= 2; m++) {
6208 GemmMicrokernelTester()
6209 .mr(2)
6210 .nr(8)
6211 .kr(2)
6212 .sr(4)
6213 .m(m)
6214 .n(n)
6215 .k(k)
6216 .iterations(1)
6217 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6218 }
6219 }
6220 }
6221 }
6222
6223 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_gt_8) {
6224 TEST_REQUIRES_ARM_NEON;
6225 for (size_t k = 9; k < 16; k++) {
6226 GemmMicrokernelTester()
6227 .mr(2)
6228 .nr(8)
6229 .kr(2)
6230 .sr(4)
6231 .m(2)
6232 .n(8)
6233 .k(k)
6234 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6235 }
6236 }
6237
6238 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_gt_8_subtile) {
6239 TEST_REQUIRES_ARM_NEON;
6240 for (size_t k = 9; k < 16; k++) {
6241 for (uint32_t n = 1; n <= 8; n++) {
6242 for (uint32_t m = 1; m <= 2; m++) {
6243 GemmMicrokernelTester()
6244 .mr(2)
6245 .nr(8)
6246 .kr(2)
6247 .sr(4)
6248 .m(m)
6249 .n(n)
6250 .k(k)
6251 .iterations(1)
6252 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6253 }
6254 }
6255 }
6256 }
6257
6258 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_div_8) {
6259 TEST_REQUIRES_ARM_NEON;
6260 for (size_t k = 16; k <= 80; k += 8) {
6261 GemmMicrokernelTester()
6262 .mr(2)
6263 .nr(8)
6264 .kr(2)
6265 .sr(4)
6266 .m(2)
6267 .n(8)
6268 .k(k)
6269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6270 }
6271 }
6272
6273 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, k_div_8_subtile) {
6274 TEST_REQUIRES_ARM_NEON;
6275 for (size_t k = 16; k <= 80; k += 8) {
6276 for (uint32_t n = 1; n <= 8; n++) {
6277 for (uint32_t m = 1; m <= 2; m++) {
6278 GemmMicrokernelTester()
6279 .mr(2)
6280 .nr(8)
6281 .kr(2)
6282 .sr(4)
6283 .m(m)
6284 .n(n)
6285 .k(k)
6286 .iterations(1)
6287 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6288 }
6289 }
6290 }
6291 }
6292
6293 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_gt_8) {
6294 TEST_REQUIRES_ARM_NEON;
6295 for (uint32_t n = 9; n < 16; n++) {
6296 for (size_t k = 1; k <= 40; k += 9) {
6297 GemmMicrokernelTester()
6298 .mr(2)
6299 .nr(8)
6300 .kr(2)
6301 .sr(4)
6302 .m(2)
6303 .n(n)
6304 .k(k)
6305 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6306 }
6307 }
6308 }
6309
6310 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_gt_8_strided_cn) {
6311 TEST_REQUIRES_ARM_NEON;
6312 for (uint32_t n = 9; n < 16; n++) {
6313 for (size_t k = 1; k <= 40; k += 9) {
6314 GemmMicrokernelTester()
6315 .mr(2)
6316 .nr(8)
6317 .kr(2)
6318 .sr(4)
6319 .m(2)
6320 .n(n)
6321 .k(k)
6322 .cn_stride(11)
6323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6324 }
6325 }
6326 }
6327
6328 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_gt_8_subtile) {
6329 TEST_REQUIRES_ARM_NEON;
6330 for (uint32_t n = 9; n < 16; n++) {
6331 for (size_t k = 1; k <= 40; k += 9) {
6332 for (uint32_t m = 1; m <= 2; m++) {
6333 GemmMicrokernelTester()
6334 .mr(2)
6335 .nr(8)
6336 .kr(2)
6337 .sr(4)
6338 .m(m)
6339 .n(n)
6340 .k(k)
6341 .iterations(1)
6342 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6343 }
6344 }
6345 }
6346 }
6347
6348 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_div_8) {
6349 TEST_REQUIRES_ARM_NEON;
6350 for (uint32_t n = 16; n <= 24; n += 8) {
6351 for (size_t k = 1; k <= 40; k += 9) {
6352 GemmMicrokernelTester()
6353 .mr(2)
6354 .nr(8)
6355 .kr(2)
6356 .sr(4)
6357 .m(2)
6358 .n(n)
6359 .k(k)
6360 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6361 }
6362 }
6363 }
6364
6365 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_div_8_strided_cn) {
6366 TEST_REQUIRES_ARM_NEON;
6367 for (uint32_t n = 16; n <= 24; n += 8) {
6368 for (size_t k = 1; k <= 40; k += 9) {
6369 GemmMicrokernelTester()
6370 .mr(2)
6371 .nr(8)
6372 .kr(2)
6373 .sr(4)
6374 .m(2)
6375 .n(n)
6376 .k(k)
6377 .cn_stride(11)
6378 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6379 }
6380 }
6381 }
6382
6383 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_div_8_subtile) {
6384 TEST_REQUIRES_ARM_NEON;
6385 for (uint32_t n = 16; n <= 24; n += 8) {
6386 for (size_t k = 1; k <= 40; k += 9) {
6387 for (uint32_t m = 1; m <= 2; m++) {
6388 GemmMicrokernelTester()
6389 .mr(2)
6390 .nr(8)
6391 .kr(2)
6392 .sr(4)
6393 .m(m)
6394 .n(n)
6395 .k(k)
6396 .iterations(1)
6397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6398 }
6399 }
6400 }
6401 }
6402
6403 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, small_kernel) {
6404 TEST_REQUIRES_ARM_NEON;
6405 for (size_t k = 1; k <= 40; k += 9) {
6406 GemmMicrokernelTester()
6407 .mr(2)
6408 .nr(8)
6409 .kr(2)
6410 .sr(4)
6411 .m(2)
6412 .n(8)
6413 .k(k)
6414 .ks(3)
6415 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6416 }
6417 }
6418
6419 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, small_kernel_subtile) {
6420 TEST_REQUIRES_ARM_NEON;
6421 for (size_t k = 1; k <= 40; k += 9) {
6422 for (uint32_t n = 1; n <= 8; n++) {
6423 for (uint32_t m = 1; m <= 2; m++) {
6424 GemmMicrokernelTester()
6425 .mr(2)
6426 .nr(8)
6427 .kr(2)
6428 .sr(4)
6429 .m(m)
6430 .n(n)
6431 .k(k)
6432 .ks(3)
6433 .iterations(1)
6434 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6435 }
6436 }
6437 }
6438 }
6439
6440 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_gt_8_small_kernel) {
6441 TEST_REQUIRES_ARM_NEON;
6442 for (uint32_t n = 9; n < 16; n++) {
6443 for (size_t k = 1; k <= 40; k += 9) {
6444 GemmMicrokernelTester()
6445 .mr(2)
6446 .nr(8)
6447 .kr(2)
6448 .sr(4)
6449 .m(2)
6450 .n(n)
6451 .k(k)
6452 .ks(3)
6453 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6454 }
6455 }
6456 }
6457
6458 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, n_div_8_small_kernel) {
6459 TEST_REQUIRES_ARM_NEON;
6460 for (uint32_t n = 16; n <= 24; n += 8) {
6461 for (size_t k = 1; k <= 40; k += 9) {
6462 GemmMicrokernelTester()
6463 .mr(2)
6464 .nr(8)
6465 .kr(2)
6466 .sr(4)
6467 .m(2)
6468 .n(n)
6469 .k(k)
6470 .ks(3)
6471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6472 }
6473 }
6474 }
6475
6476 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, strided_cm_subtile) {
6477 TEST_REQUIRES_ARM_NEON;
6478 for (size_t k = 1; k <= 40; k += 9) {
6479 for (uint32_t n = 1; n <= 8; n++) {
6480 for (uint32_t m = 1; m <= 2; m++) {
6481 GemmMicrokernelTester()
6482 .mr(2)
6483 .nr(8)
6484 .kr(2)
6485 .sr(4)
6486 .m(m)
6487 .n(n)
6488 .k(k)
6489 .cm_stride(11)
6490 .iterations(1)
6491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6492 }
6493 }
6494 }
6495 }
6496
6497 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, a_offset) {
6498 TEST_REQUIRES_ARM_NEON;
6499 for (size_t k = 1; k <= 40; k += 9) {
6500 GemmMicrokernelTester()
6501 .mr(2)
6502 .nr(8)
6503 .kr(2)
6504 .sr(4)
6505 .m(2)
6506 .n(8)
6507 .k(k)
6508 .ks(3)
6509 .a_offset(83)
6510 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6511 }
6512 }
6513
6514 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, zero) {
6515 TEST_REQUIRES_ARM_NEON;
6516 for (size_t k = 1; k <= 40; k += 9) {
6517 for (uint32_t mz = 0; mz < 2; mz++) {
6518 GemmMicrokernelTester()
6519 .mr(2)
6520 .nr(8)
6521 .kr(2)
6522 .sr(4)
6523 .m(2)
6524 .n(8)
6525 .k(k)
6526 .ks(3)
6527 .a_offset(83)
6528 .zero_index(mz)
6529 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6530 }
6531 }
6532 }
6533
6534 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, qmin) {
6535 TEST_REQUIRES_ARM_NEON;
6536 GemmMicrokernelTester()
6537 .mr(2)
6538 .nr(8)
6539 .kr(2)
6540 .sr(4)
6541 .m(2)
6542 .n(8)
6543 .k(8)
6544 .qmin(128)
6545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6546 }
6547
6548 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, qmax) {
6549 TEST_REQUIRES_ARM_NEON;
6550 GemmMicrokernelTester()
6551 .mr(2)
6552 .nr(8)
6553 .kr(2)
6554 .sr(4)
6555 .m(2)
6556 .n(8)
6557 .k(8)
6558 .qmax(128)
6559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6560 }
6561
6562 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C2S4__NEON_MULL, strided_cm) {
6563 TEST_REQUIRES_ARM_NEON;
6564 GemmMicrokernelTester()
6565 .mr(2)
6566 .nr(8)
6567 .kr(2)
6568 .sr(4)
6569 .m(2)
6570 .n(8)
6571 .k(8)
6572 .cm_stride(11)
6573 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6574 }
6575#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
6576
6577
6578#if XNN_ARCH_ARM || XNN_ARCH_ARM64
6579 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_eq_8) {
6580 TEST_REQUIRES_ARM_NEON;
6581 GemmMicrokernelTester()
6582 .mr(1)
6583 .nr(16)
6584 .kr(2)
6585 .sr(4)
6586 .m(1)
6587 .n(16)
6588 .k(8)
6589 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6590 }
6591
6592 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, strided_cn) {
6593 TEST_REQUIRES_ARM_NEON;
6594 GemmMicrokernelTester()
6595 .mr(1)
6596 .nr(16)
6597 .kr(2)
6598 .sr(4)
6599 .m(1)
6600 .n(16)
6601 .k(8)
6602 .cn_stride(19)
6603 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6604 }
6605
6606 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_eq_8_subtile) {
6607 TEST_REQUIRES_ARM_NEON;
6608 for (uint32_t n = 1; n <= 16; n++) {
6609 for (uint32_t m = 1; m <= 1; m++) {
6610 GemmMicrokernelTester()
6611 .mr(1)
6612 .nr(16)
6613 .kr(2)
6614 .sr(4)
6615 .m(m)
6616 .n(n)
6617 .k(8)
6618 .iterations(1)
6619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6620 }
6621 }
6622 }
6623
6624 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_eq_8_subtile_m) {
6625 TEST_REQUIRES_ARM_NEON;
6626 for (uint32_t m = 1; m <= 1; m++) {
6627 GemmMicrokernelTester()
6628 .mr(1)
6629 .nr(16)
6630 .kr(2)
6631 .sr(4)
6632 .m(m)
6633 .n(16)
6634 .k(8)
6635 .iterations(1)
6636 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6637 }
6638 }
6639
6640 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_eq_8_subtile_n) {
6641 TEST_REQUIRES_ARM_NEON;
6642 for (uint32_t n = 1; n <= 16; n++) {
6643 GemmMicrokernelTester()
6644 .mr(1)
6645 .nr(16)
6646 .kr(2)
6647 .sr(4)
6648 .m(1)
6649 .n(n)
6650 .k(8)
6651 .iterations(1)
6652 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6653 }
6654 }
6655
6656 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_lt_8) {
6657 TEST_REQUIRES_ARM_NEON;
6658 for (size_t k = 1; k < 8; k++) {
6659 GemmMicrokernelTester()
6660 .mr(1)
6661 .nr(16)
6662 .kr(2)
6663 .sr(4)
6664 .m(1)
6665 .n(16)
6666 .k(k)
6667 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6668 }
6669 }
6670
6671 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_lt_8_subtile) {
6672 TEST_REQUIRES_ARM_NEON;
6673 for (size_t k = 1; k < 8; k++) {
6674 for (uint32_t n = 1; n <= 16; n++) {
6675 for (uint32_t m = 1; m <= 1; m++) {
6676 GemmMicrokernelTester()
6677 .mr(1)
6678 .nr(16)
6679 .kr(2)
6680 .sr(4)
6681 .m(m)
6682 .n(n)
6683 .k(k)
6684 .iterations(1)
6685 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6686 }
6687 }
6688 }
6689 }
6690
6691 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_gt_8) {
6692 TEST_REQUIRES_ARM_NEON;
6693 for (size_t k = 9; k < 16; k++) {
6694 GemmMicrokernelTester()
6695 .mr(1)
6696 .nr(16)
6697 .kr(2)
6698 .sr(4)
6699 .m(1)
6700 .n(16)
6701 .k(k)
6702 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6703 }
6704 }
6705
6706 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_gt_8_subtile) {
6707 TEST_REQUIRES_ARM_NEON;
6708 for (size_t k = 9; k < 16; k++) {
6709 for (uint32_t n = 1; n <= 16; n++) {
6710 for (uint32_t m = 1; m <= 1; m++) {
6711 GemmMicrokernelTester()
6712 .mr(1)
6713 .nr(16)
6714 .kr(2)
6715 .sr(4)
6716 .m(m)
6717 .n(n)
6718 .k(k)
6719 .iterations(1)
6720 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6721 }
6722 }
6723 }
6724 }
6725
6726 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_div_8) {
6727 TEST_REQUIRES_ARM_NEON;
6728 for (size_t k = 16; k <= 80; k += 8) {
6729 GemmMicrokernelTester()
6730 .mr(1)
6731 .nr(16)
6732 .kr(2)
6733 .sr(4)
6734 .m(1)
6735 .n(16)
6736 .k(k)
6737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6738 }
6739 }
6740
6741 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, k_div_8_subtile) {
6742 TEST_REQUIRES_ARM_NEON;
6743 for (size_t k = 16; k <= 80; k += 8) {
6744 for (uint32_t n = 1; n <= 16; n++) {
6745 for (uint32_t m = 1; m <= 1; m++) {
6746 GemmMicrokernelTester()
6747 .mr(1)
6748 .nr(16)
6749 .kr(2)
6750 .sr(4)
6751 .m(m)
6752 .n(n)
6753 .k(k)
6754 .iterations(1)
6755 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6756 }
6757 }
6758 }
6759 }
6760
6761 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_gt_16) {
6762 TEST_REQUIRES_ARM_NEON;
6763 for (uint32_t n = 17; n < 32; n++) {
6764 for (size_t k = 1; k <= 40; k += 9) {
6765 GemmMicrokernelTester()
6766 .mr(1)
6767 .nr(16)
6768 .kr(2)
6769 .sr(4)
6770 .m(1)
6771 .n(n)
6772 .k(k)
6773 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6774 }
6775 }
6776 }
6777
6778 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_gt_16_strided_cn) {
6779 TEST_REQUIRES_ARM_NEON;
6780 for (uint32_t n = 17; n < 32; n++) {
6781 for (size_t k = 1; k <= 40; k += 9) {
6782 GemmMicrokernelTester()
6783 .mr(1)
6784 .nr(16)
6785 .kr(2)
6786 .sr(4)
6787 .m(1)
6788 .n(n)
6789 .k(k)
6790 .cn_stride(19)
6791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6792 }
6793 }
6794 }
6795
6796 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_gt_16_subtile) {
6797 TEST_REQUIRES_ARM_NEON;
6798 for (uint32_t n = 17; n < 32; n++) {
6799 for (size_t k = 1; k <= 40; k += 9) {
6800 for (uint32_t m = 1; m <= 1; m++) {
6801 GemmMicrokernelTester()
6802 .mr(1)
6803 .nr(16)
6804 .kr(2)
6805 .sr(4)
6806 .m(m)
6807 .n(n)
6808 .k(k)
6809 .iterations(1)
6810 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6811 }
6812 }
6813 }
6814 }
6815
6816 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_div_16) {
6817 TEST_REQUIRES_ARM_NEON;
6818 for (uint32_t n = 32; n <= 48; n += 16) {
6819 for (size_t k = 1; k <= 40; k += 9) {
6820 GemmMicrokernelTester()
6821 .mr(1)
6822 .nr(16)
6823 .kr(2)
6824 .sr(4)
6825 .m(1)
6826 .n(n)
6827 .k(k)
6828 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6829 }
6830 }
6831 }
6832
6833 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_div_16_strided_cn) {
6834 TEST_REQUIRES_ARM_NEON;
6835 for (uint32_t n = 32; n <= 48; n += 16) {
6836 for (size_t k = 1; k <= 40; k += 9) {
6837 GemmMicrokernelTester()
6838 .mr(1)
6839 .nr(16)
6840 .kr(2)
6841 .sr(4)
6842 .m(1)
6843 .n(n)
6844 .k(k)
6845 .cn_stride(19)
6846 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6847 }
6848 }
6849 }
6850
6851 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_div_16_subtile) {
6852 TEST_REQUIRES_ARM_NEON;
6853 for (uint32_t n = 32; n <= 48; n += 16) {
6854 for (size_t k = 1; k <= 40; k += 9) {
6855 for (uint32_t m = 1; m <= 1; m++) {
6856 GemmMicrokernelTester()
6857 .mr(1)
6858 .nr(16)
6859 .kr(2)
6860 .sr(4)
6861 .m(m)
6862 .n(n)
6863 .k(k)
6864 .iterations(1)
6865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6866 }
6867 }
6868 }
6869 }
6870
6871 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, small_kernel) {
6872 TEST_REQUIRES_ARM_NEON;
6873 for (size_t k = 1; k <= 40; k += 9) {
6874 GemmMicrokernelTester()
6875 .mr(1)
6876 .nr(16)
6877 .kr(2)
6878 .sr(4)
6879 .m(1)
6880 .n(16)
6881 .k(k)
6882 .ks(3)
6883 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6884 }
6885 }
6886
6887 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, small_kernel_subtile) {
6888 TEST_REQUIRES_ARM_NEON;
6889 for (size_t k = 1; k <= 40; k += 9) {
6890 for (uint32_t n = 1; n <= 16; n++) {
6891 for (uint32_t m = 1; m <= 1; m++) {
6892 GemmMicrokernelTester()
6893 .mr(1)
6894 .nr(16)
6895 .kr(2)
6896 .sr(4)
6897 .m(m)
6898 .n(n)
6899 .k(k)
6900 .ks(3)
6901 .iterations(1)
6902 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6903 }
6904 }
6905 }
6906 }
6907
6908 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_gt_16_small_kernel) {
6909 TEST_REQUIRES_ARM_NEON;
6910 for (uint32_t n = 17; n < 32; n++) {
6911 for (size_t k = 1; k <= 40; k += 9) {
6912 GemmMicrokernelTester()
6913 .mr(1)
6914 .nr(16)
6915 .kr(2)
6916 .sr(4)
6917 .m(1)
6918 .n(n)
6919 .k(k)
6920 .ks(3)
6921 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6922 }
6923 }
6924 }
6925
6926 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, n_div_16_small_kernel) {
6927 TEST_REQUIRES_ARM_NEON;
6928 for (uint32_t n = 32; n <= 48; n += 16) {
6929 for (size_t k = 1; k <= 40; k += 9) {
6930 GemmMicrokernelTester()
6931 .mr(1)
6932 .nr(16)
6933 .kr(2)
6934 .sr(4)
6935 .m(1)
6936 .n(n)
6937 .k(k)
6938 .ks(3)
6939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6940 }
6941 }
6942 }
6943
6944 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, strided_cm_subtile) {
6945 TEST_REQUIRES_ARM_NEON;
6946 for (size_t k = 1; k <= 40; k += 9) {
6947 for (uint32_t n = 1; n <= 16; n++) {
6948 for (uint32_t m = 1; m <= 1; m++) {
6949 GemmMicrokernelTester()
6950 .mr(1)
6951 .nr(16)
6952 .kr(2)
6953 .sr(4)
6954 .m(m)
6955 .n(n)
6956 .k(k)
6957 .cm_stride(19)
6958 .iterations(1)
6959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6960 }
6961 }
6962 }
6963 }
6964
6965 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, a_offset) {
6966 TEST_REQUIRES_ARM_NEON;
6967 for (size_t k = 1; k <= 40; k += 9) {
6968 GemmMicrokernelTester()
6969 .mr(1)
6970 .nr(16)
6971 .kr(2)
6972 .sr(4)
6973 .m(1)
6974 .n(16)
6975 .k(k)
6976 .ks(3)
6977 .a_offset(43)
6978 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6979 }
6980 }
6981
6982 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, zero) {
6983 TEST_REQUIRES_ARM_NEON;
6984 for (size_t k = 1; k <= 40; k += 9) {
6985 for (uint32_t mz = 0; mz < 1; mz++) {
6986 GemmMicrokernelTester()
6987 .mr(1)
6988 .nr(16)
6989 .kr(2)
6990 .sr(4)
6991 .m(1)
6992 .n(16)
6993 .k(k)
6994 .ks(3)
6995 .a_offset(43)
6996 .zero_index(mz)
6997 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
6998 }
6999 }
7000 }
7001
7002 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, qmin) {
7003 TEST_REQUIRES_ARM_NEON;
7004 GemmMicrokernelTester()
7005 .mr(1)
7006 .nr(16)
7007 .kr(2)
7008 .sr(4)
7009 .m(1)
7010 .n(16)
7011 .k(8)
7012 .qmin(128)
7013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7014 }
7015
7016 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, qmax) {
7017 TEST_REQUIRES_ARM_NEON;
7018 GemmMicrokernelTester()
7019 .mr(1)
7020 .nr(16)
7021 .kr(2)
7022 .sr(4)
7023 .m(1)
7024 .n(16)
7025 .k(8)
7026 .qmax(128)
7027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7028 }
7029
7030 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2S4__NEON_MULL, strided_cm) {
7031 TEST_REQUIRES_ARM_NEON;
7032 GemmMicrokernelTester()
7033 .mr(1)
7034 .nr(16)
7035 .kr(2)
7036 .sr(4)
7037 .m(1)
7038 .n(16)
7039 .k(8)
7040 .cm_stride(19)
7041 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2s4__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7042 }
7043#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
7044
7045
7046#if XNN_ARCH_ARM || XNN_ARCH_ARM64
7047 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_eq_16) {
7048 TEST_REQUIRES_ARM_NEON;
7049 GemmMicrokernelTester()
7050 .mr(1)
7051 .nr(8)
7052 .kr(2)
7053 .sr(4)
7054 .m(1)
7055 .n(8)
7056 .k(16)
7057 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7058 }
7059
7060 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, strided_cn) {
7061 TEST_REQUIRES_ARM_NEON;
7062 GemmMicrokernelTester()
7063 .mr(1)
7064 .nr(8)
7065 .kr(2)
7066 .sr(4)
7067 .m(1)
7068 .n(8)
7069 .k(16)
7070 .cn_stride(11)
7071 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7072 }
7073
7074 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_eq_16_subtile) {
7075 TEST_REQUIRES_ARM_NEON;
7076 for (uint32_t n = 1; n <= 8; n++) {
7077 for (uint32_t m = 1; m <= 1; m++) {
7078 GemmMicrokernelTester()
7079 .mr(1)
7080 .nr(8)
7081 .kr(2)
7082 .sr(4)
7083 .m(m)
7084 .n(n)
7085 .k(16)
7086 .iterations(1)
7087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7088 }
7089 }
7090 }
7091
7092 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_eq_16_subtile_m) {
7093 TEST_REQUIRES_ARM_NEON;
7094 for (uint32_t m = 1; m <= 1; m++) {
7095 GemmMicrokernelTester()
7096 .mr(1)
7097 .nr(8)
7098 .kr(2)
7099 .sr(4)
7100 .m(m)
7101 .n(8)
7102 .k(16)
7103 .iterations(1)
7104 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7105 }
7106 }
7107
7108 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_eq_16_subtile_n) {
7109 TEST_REQUIRES_ARM_NEON;
7110 for (uint32_t n = 1; n <= 8; n++) {
7111 GemmMicrokernelTester()
7112 .mr(1)
7113 .nr(8)
7114 .kr(2)
7115 .sr(4)
7116 .m(1)
7117 .n(n)
7118 .k(16)
7119 .iterations(1)
7120 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7121 }
7122 }
7123
7124 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_lt_16) {
7125 TEST_REQUIRES_ARM_NEON;
7126 for (size_t k = 1; k < 16; k++) {
7127 GemmMicrokernelTester()
7128 .mr(1)
7129 .nr(8)
7130 .kr(2)
7131 .sr(4)
7132 .m(1)
7133 .n(8)
7134 .k(k)
7135 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7136 }
7137 }
7138
7139 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_lt_16_subtile) {
7140 TEST_REQUIRES_ARM_NEON;
7141 for (size_t k = 1; k < 16; k++) {
7142 for (uint32_t n = 1; n <= 8; n++) {
7143 for (uint32_t m = 1; m <= 1; m++) {
7144 GemmMicrokernelTester()
7145 .mr(1)
7146 .nr(8)
7147 .kr(2)
7148 .sr(4)
7149 .m(m)
7150 .n(n)
7151 .k(k)
7152 .iterations(1)
7153 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7154 }
7155 }
7156 }
7157 }
7158
7159 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_gt_16) {
7160 TEST_REQUIRES_ARM_NEON;
7161 for (size_t k = 17; k < 32; k++) {
7162 GemmMicrokernelTester()
7163 .mr(1)
7164 .nr(8)
7165 .kr(2)
7166 .sr(4)
7167 .m(1)
7168 .n(8)
7169 .k(k)
7170 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7171 }
7172 }
7173
7174 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_gt_16_subtile) {
7175 TEST_REQUIRES_ARM_NEON;
7176 for (size_t k = 17; k < 32; k++) {
7177 for (uint32_t n = 1; n <= 8; n++) {
7178 for (uint32_t m = 1; m <= 1; m++) {
7179 GemmMicrokernelTester()
7180 .mr(1)
7181 .nr(8)
7182 .kr(2)
7183 .sr(4)
7184 .m(m)
7185 .n(n)
7186 .k(k)
7187 .iterations(1)
7188 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7189 }
7190 }
7191 }
7192 }
7193
7194 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_div_16) {
7195 TEST_REQUIRES_ARM_NEON;
7196 for (size_t k = 32; k <= 160; k += 16) {
7197 GemmMicrokernelTester()
7198 .mr(1)
7199 .nr(8)
7200 .kr(2)
7201 .sr(4)
7202 .m(1)
7203 .n(8)
7204 .k(k)
7205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7206 }
7207 }
7208
7209 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, k_div_16_subtile) {
7210 TEST_REQUIRES_ARM_NEON;
7211 for (size_t k = 32; k <= 160; k += 16) {
7212 for (uint32_t n = 1; n <= 8; n++) {
7213 for (uint32_t m = 1; m <= 1; m++) {
7214 GemmMicrokernelTester()
7215 .mr(1)
7216 .nr(8)
7217 .kr(2)
7218 .sr(4)
7219 .m(m)
7220 .n(n)
7221 .k(k)
7222 .iterations(1)
7223 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7224 }
7225 }
7226 }
7227 }
7228
7229 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_gt_8) {
7230 TEST_REQUIRES_ARM_NEON;
7231 for (uint32_t n = 9; n < 16; n++) {
7232 for (size_t k = 1; k <= 80; k += 17) {
7233 GemmMicrokernelTester()
7234 .mr(1)
7235 .nr(8)
7236 .kr(2)
7237 .sr(4)
7238 .m(1)
7239 .n(n)
7240 .k(k)
7241 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7242 }
7243 }
7244 }
7245
7246 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_gt_8_strided_cn) {
7247 TEST_REQUIRES_ARM_NEON;
7248 for (uint32_t n = 9; n < 16; n++) {
7249 for (size_t k = 1; k <= 80; k += 17) {
7250 GemmMicrokernelTester()
7251 .mr(1)
7252 .nr(8)
7253 .kr(2)
7254 .sr(4)
7255 .m(1)
7256 .n(n)
7257 .k(k)
7258 .cn_stride(11)
7259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7260 }
7261 }
7262 }
7263
7264 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_gt_8_subtile) {
7265 TEST_REQUIRES_ARM_NEON;
7266 for (uint32_t n = 9; n < 16; n++) {
7267 for (size_t k = 1; k <= 80; k += 17) {
7268 for (uint32_t m = 1; m <= 1; m++) {
7269 GemmMicrokernelTester()
7270 .mr(1)
7271 .nr(8)
7272 .kr(2)
7273 .sr(4)
7274 .m(m)
7275 .n(n)
7276 .k(k)
7277 .iterations(1)
7278 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7279 }
7280 }
7281 }
7282 }
7283
7284 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_div_8) {
7285 TEST_REQUIRES_ARM_NEON;
7286 for (uint32_t n = 16; n <= 24; n += 8) {
7287 for (size_t k = 1; k <= 80; k += 17) {
7288 GemmMicrokernelTester()
7289 .mr(1)
7290 .nr(8)
7291 .kr(2)
7292 .sr(4)
7293 .m(1)
7294 .n(n)
7295 .k(k)
7296 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7297 }
7298 }
7299 }
7300
7301 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_div_8_strided_cn) {
7302 TEST_REQUIRES_ARM_NEON;
7303 for (uint32_t n = 16; n <= 24; n += 8) {
7304 for (size_t k = 1; k <= 80; k += 17) {
7305 GemmMicrokernelTester()
7306 .mr(1)
7307 .nr(8)
7308 .kr(2)
7309 .sr(4)
7310 .m(1)
7311 .n(n)
7312 .k(k)
7313 .cn_stride(11)
7314 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7315 }
7316 }
7317 }
7318
7319 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_div_8_subtile) {
7320 TEST_REQUIRES_ARM_NEON;
7321 for (uint32_t n = 16; n <= 24; n += 8) {
7322 for (size_t k = 1; k <= 80; k += 17) {
7323 for (uint32_t m = 1; m <= 1; m++) {
7324 GemmMicrokernelTester()
7325 .mr(1)
7326 .nr(8)
7327 .kr(2)
7328 .sr(4)
7329 .m(m)
7330 .n(n)
7331 .k(k)
7332 .iterations(1)
7333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7334 }
7335 }
7336 }
7337 }
7338
7339 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, small_kernel) {
7340 TEST_REQUIRES_ARM_NEON;
7341 for (size_t k = 1; k <= 80; k += 17) {
7342 GemmMicrokernelTester()
7343 .mr(1)
7344 .nr(8)
7345 .kr(2)
7346 .sr(4)
7347 .m(1)
7348 .n(8)
7349 .k(k)
7350 .ks(3)
7351 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7352 }
7353 }
7354
7355 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, small_kernel_subtile) {
7356 TEST_REQUIRES_ARM_NEON;
7357 for (size_t k = 1; k <= 80; k += 17) {
7358 for (uint32_t n = 1; n <= 8; n++) {
7359 for (uint32_t m = 1; m <= 1; m++) {
7360 GemmMicrokernelTester()
7361 .mr(1)
7362 .nr(8)
7363 .kr(2)
7364 .sr(4)
7365 .m(m)
7366 .n(n)
7367 .k(k)
7368 .ks(3)
7369 .iterations(1)
7370 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7371 }
7372 }
7373 }
7374 }
7375
7376 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_gt_8_small_kernel) {
7377 TEST_REQUIRES_ARM_NEON;
7378 for (uint32_t n = 9; n < 16; n++) {
7379 for (size_t k = 1; k <= 80; k += 17) {
7380 GemmMicrokernelTester()
7381 .mr(1)
7382 .nr(8)
7383 .kr(2)
7384 .sr(4)
7385 .m(1)
7386 .n(n)
7387 .k(k)
7388 .ks(3)
7389 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7390 }
7391 }
7392 }
7393
7394 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, n_div_8_small_kernel) {
7395 TEST_REQUIRES_ARM_NEON;
7396 for (uint32_t n = 16; n <= 24; n += 8) {
7397 for (size_t k = 1; k <= 80; k += 17) {
7398 GemmMicrokernelTester()
7399 .mr(1)
7400 .nr(8)
7401 .kr(2)
7402 .sr(4)
7403 .m(1)
7404 .n(n)
7405 .k(k)
7406 .ks(3)
7407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7408 }
7409 }
7410 }
7411
7412 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, strided_cm_subtile) {
7413 TEST_REQUIRES_ARM_NEON;
7414 for (size_t k = 1; k <= 80; k += 17) {
7415 for (uint32_t n = 1; n <= 8; n++) {
7416 for (uint32_t m = 1; m <= 1; m++) {
7417 GemmMicrokernelTester()
7418 .mr(1)
7419 .nr(8)
7420 .kr(2)
7421 .sr(4)
7422 .m(m)
7423 .n(n)
7424 .k(k)
7425 .cm_stride(11)
7426 .iterations(1)
7427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7428 }
7429 }
7430 }
7431 }
7432
7433 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, a_offset) {
7434 TEST_REQUIRES_ARM_NEON;
7435 for (size_t k = 1; k <= 80; k += 17) {
7436 GemmMicrokernelTester()
7437 .mr(1)
7438 .nr(8)
7439 .kr(2)
7440 .sr(4)
7441 .m(1)
7442 .n(8)
7443 .k(k)
7444 .ks(3)
7445 .a_offset(83)
7446 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7447 }
7448 }
7449
7450 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, zero) {
7451 TEST_REQUIRES_ARM_NEON;
7452 for (size_t k = 1; k <= 80; k += 17) {
7453 for (uint32_t mz = 0; mz < 1; mz++) {
7454 GemmMicrokernelTester()
7455 .mr(1)
7456 .nr(8)
7457 .kr(2)
7458 .sr(4)
7459 .m(1)
7460 .n(8)
7461 .k(k)
7462 .ks(3)
7463 .a_offset(83)
7464 .zero_index(mz)
7465 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7466 }
7467 }
7468 }
7469
7470 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, qmin) {
7471 TEST_REQUIRES_ARM_NEON;
7472 GemmMicrokernelTester()
7473 .mr(1)
7474 .nr(8)
7475 .kr(2)
7476 .sr(4)
7477 .m(1)
7478 .n(8)
7479 .k(16)
7480 .qmin(128)
7481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7482 }
7483
7484 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, qmax) {
7485 TEST_REQUIRES_ARM_NEON;
7486 GemmMicrokernelTester()
7487 .mr(1)
7488 .nr(8)
7489 .kr(2)
7490 .sr(4)
7491 .m(1)
7492 .n(8)
7493 .k(16)
7494 .qmax(128)
7495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7496 }
7497
7498 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2S4__NEON_MLAL, strided_cm) {
7499 TEST_REQUIRES_ARM_NEON;
7500 GemmMicrokernelTester()
7501 .mr(1)
7502 .nr(8)
7503 .kr(2)
7504 .sr(4)
7505 .m(1)
7506 .n(8)
7507 .k(16)
7508 .cm_stride(11)
7509 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7510 }
7511#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
7512
7513
7514#if XNN_ARCH_ARM || XNN_ARCH_ARM64
7515 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_eq_16) {
7516 TEST_REQUIRES_ARM_NEON;
7517 GemmMicrokernelTester()
7518 .mr(4)
7519 .nr(16)
7520 .kr(2)
7521 .sr(4)
7522 .m(4)
7523 .n(16)
7524 .k(16)
7525 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7526 }
7527
7528 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, strided_cn) {
7529 TEST_REQUIRES_ARM_NEON;
7530 GemmMicrokernelTester()
7531 .mr(4)
7532 .nr(16)
7533 .kr(2)
7534 .sr(4)
7535 .m(4)
7536 .n(16)
7537 .k(16)
7538 .cn_stride(19)
7539 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7540 }
7541
7542 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_eq_16_subtile) {
7543 TEST_REQUIRES_ARM_NEON;
7544 for (uint32_t n = 1; n <= 16; n++) {
7545 for (uint32_t m = 1; m <= 4; m++) {
7546 GemmMicrokernelTester()
7547 .mr(4)
7548 .nr(16)
7549 .kr(2)
7550 .sr(4)
7551 .m(m)
7552 .n(n)
7553 .k(16)
7554 .iterations(1)
7555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7556 }
7557 }
7558 }
7559
7560 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_eq_16_subtile_m) {
7561 TEST_REQUIRES_ARM_NEON;
7562 for (uint32_t m = 1; m <= 4; m++) {
7563 GemmMicrokernelTester()
7564 .mr(4)
7565 .nr(16)
7566 .kr(2)
7567 .sr(4)
7568 .m(m)
7569 .n(16)
7570 .k(16)
7571 .iterations(1)
7572 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7573 }
7574 }
7575
7576 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_eq_16_subtile_n) {
7577 TEST_REQUIRES_ARM_NEON;
7578 for (uint32_t n = 1; n <= 16; n++) {
7579 GemmMicrokernelTester()
7580 .mr(4)
7581 .nr(16)
7582 .kr(2)
7583 .sr(4)
7584 .m(4)
7585 .n(n)
7586 .k(16)
7587 .iterations(1)
7588 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7589 }
7590 }
7591
7592 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_lt_16) {
7593 TEST_REQUIRES_ARM_NEON;
7594 for (size_t k = 1; k < 16; k++) {
7595 GemmMicrokernelTester()
7596 .mr(4)
7597 .nr(16)
7598 .kr(2)
7599 .sr(4)
7600 .m(4)
7601 .n(16)
7602 .k(k)
7603 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7604 }
7605 }
7606
7607 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_lt_16_subtile) {
7608 TEST_REQUIRES_ARM_NEON;
7609 for (size_t k = 1; k < 16; k++) {
7610 for (uint32_t n = 1; n <= 16; n++) {
7611 for (uint32_t m = 1; m <= 4; m++) {
7612 GemmMicrokernelTester()
7613 .mr(4)
7614 .nr(16)
7615 .kr(2)
7616 .sr(4)
7617 .m(m)
7618 .n(n)
7619 .k(k)
7620 .iterations(1)
7621 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7622 }
7623 }
7624 }
7625 }
7626
7627 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_gt_16) {
7628 TEST_REQUIRES_ARM_NEON;
7629 for (size_t k = 17; k < 32; k++) {
7630 GemmMicrokernelTester()
7631 .mr(4)
7632 .nr(16)
7633 .kr(2)
7634 .sr(4)
7635 .m(4)
7636 .n(16)
7637 .k(k)
7638 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7639 }
7640 }
7641
7642 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_gt_16_subtile) {
7643 TEST_REQUIRES_ARM_NEON;
7644 for (size_t k = 17; k < 32; k++) {
7645 for (uint32_t n = 1; n <= 16; n++) {
7646 for (uint32_t m = 1; m <= 4; m++) {
7647 GemmMicrokernelTester()
7648 .mr(4)
7649 .nr(16)
7650 .kr(2)
7651 .sr(4)
7652 .m(m)
7653 .n(n)
7654 .k(k)
7655 .iterations(1)
7656 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7657 }
7658 }
7659 }
7660 }
7661
7662 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_div_16) {
7663 TEST_REQUIRES_ARM_NEON;
7664 for (size_t k = 32; k <= 160; k += 16) {
7665 GemmMicrokernelTester()
7666 .mr(4)
7667 .nr(16)
7668 .kr(2)
7669 .sr(4)
7670 .m(4)
7671 .n(16)
7672 .k(k)
7673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7674 }
7675 }
7676
7677 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, k_div_16_subtile) {
7678 TEST_REQUIRES_ARM_NEON;
7679 for (size_t k = 32; k <= 160; k += 16) {
7680 for (uint32_t n = 1; n <= 16; n++) {
7681 for (uint32_t m = 1; m <= 4; m++) {
7682 GemmMicrokernelTester()
7683 .mr(4)
7684 .nr(16)
7685 .kr(2)
7686 .sr(4)
7687 .m(m)
7688 .n(n)
7689 .k(k)
7690 .iterations(1)
7691 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7692 }
7693 }
7694 }
7695 }
7696
7697 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_gt_16) {
7698 TEST_REQUIRES_ARM_NEON;
7699 for (uint32_t n = 17; n < 32; n++) {
7700 for (size_t k = 1; k <= 80; k += 17) {
7701 GemmMicrokernelTester()
7702 .mr(4)
7703 .nr(16)
7704 .kr(2)
7705 .sr(4)
7706 .m(4)
7707 .n(n)
7708 .k(k)
7709 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7710 }
7711 }
7712 }
7713
7714 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_gt_16_strided_cn) {
7715 TEST_REQUIRES_ARM_NEON;
7716 for (uint32_t n = 17; n < 32; n++) {
7717 for (size_t k = 1; k <= 80; k += 17) {
7718 GemmMicrokernelTester()
7719 .mr(4)
7720 .nr(16)
7721 .kr(2)
7722 .sr(4)
7723 .m(4)
7724 .n(n)
7725 .k(k)
7726 .cn_stride(19)
7727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7728 }
7729 }
7730 }
7731
7732 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_gt_16_subtile) {
7733 TEST_REQUIRES_ARM_NEON;
7734 for (uint32_t n = 17; n < 32; n++) {
7735 for (size_t k = 1; k <= 80; k += 17) {
7736 for (uint32_t m = 1; m <= 4; m++) {
7737 GemmMicrokernelTester()
7738 .mr(4)
7739 .nr(16)
7740 .kr(2)
7741 .sr(4)
7742 .m(m)
7743 .n(n)
7744 .k(k)
7745 .iterations(1)
7746 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7747 }
7748 }
7749 }
7750 }
7751
7752 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_div_16) {
7753 TEST_REQUIRES_ARM_NEON;
7754 for (uint32_t n = 32; n <= 48; n += 16) {
7755 for (size_t k = 1; k <= 80; k += 17) {
7756 GemmMicrokernelTester()
7757 .mr(4)
7758 .nr(16)
7759 .kr(2)
7760 .sr(4)
7761 .m(4)
7762 .n(n)
7763 .k(k)
7764 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7765 }
7766 }
7767 }
7768
7769 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_div_16_strided_cn) {
7770 TEST_REQUIRES_ARM_NEON;
7771 for (uint32_t n = 32; n <= 48; n += 16) {
7772 for (size_t k = 1; k <= 80; k += 17) {
7773 GemmMicrokernelTester()
7774 .mr(4)
7775 .nr(16)
7776 .kr(2)
7777 .sr(4)
7778 .m(4)
7779 .n(n)
7780 .k(k)
7781 .cn_stride(19)
7782 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7783 }
7784 }
7785 }
7786
7787 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_div_16_subtile) {
7788 TEST_REQUIRES_ARM_NEON;
7789 for (uint32_t n = 32; n <= 48; n += 16) {
7790 for (size_t k = 1; k <= 80; k += 17) {
7791 for (uint32_t m = 1; m <= 4; m++) {
7792 GemmMicrokernelTester()
7793 .mr(4)
7794 .nr(16)
7795 .kr(2)
7796 .sr(4)
7797 .m(m)
7798 .n(n)
7799 .k(k)
7800 .iterations(1)
7801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7802 }
7803 }
7804 }
7805 }
7806
7807 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, small_kernel) {
7808 TEST_REQUIRES_ARM_NEON;
7809 for (size_t k = 1; k <= 80; k += 17) {
7810 GemmMicrokernelTester()
7811 .mr(4)
7812 .nr(16)
7813 .kr(2)
7814 .sr(4)
7815 .m(4)
7816 .n(16)
7817 .k(k)
7818 .ks(3)
7819 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7820 }
7821 }
7822
7823 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, small_kernel_subtile) {
7824 TEST_REQUIRES_ARM_NEON;
7825 for (size_t k = 1; k <= 80; k += 17) {
7826 for (uint32_t n = 1; n <= 16; n++) {
7827 for (uint32_t m = 1; m <= 4; m++) {
7828 GemmMicrokernelTester()
7829 .mr(4)
7830 .nr(16)
7831 .kr(2)
7832 .sr(4)
7833 .m(m)
7834 .n(n)
7835 .k(k)
7836 .ks(3)
7837 .iterations(1)
7838 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7839 }
7840 }
7841 }
7842 }
7843
7844 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_gt_16_small_kernel) {
7845 TEST_REQUIRES_ARM_NEON;
7846 for (uint32_t n = 17; n < 32; n++) {
7847 for (size_t k = 1; k <= 80; k += 17) {
7848 GemmMicrokernelTester()
7849 .mr(4)
7850 .nr(16)
7851 .kr(2)
7852 .sr(4)
7853 .m(4)
7854 .n(n)
7855 .k(k)
7856 .ks(3)
7857 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7858 }
7859 }
7860 }
7861
7862 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, n_div_16_small_kernel) {
7863 TEST_REQUIRES_ARM_NEON;
7864 for (uint32_t n = 32; n <= 48; n += 16) {
7865 for (size_t k = 1; k <= 80; k += 17) {
7866 GemmMicrokernelTester()
7867 .mr(4)
7868 .nr(16)
7869 .kr(2)
7870 .sr(4)
7871 .m(4)
7872 .n(n)
7873 .k(k)
7874 .ks(3)
7875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7876 }
7877 }
7878 }
7879
7880 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, strided_cm_subtile) {
7881 TEST_REQUIRES_ARM_NEON;
7882 for (size_t k = 1; k <= 80; k += 17) {
7883 for (uint32_t n = 1; n <= 16; n++) {
7884 for (uint32_t m = 1; m <= 4; m++) {
7885 GemmMicrokernelTester()
7886 .mr(4)
7887 .nr(16)
7888 .kr(2)
7889 .sr(4)
7890 .m(m)
7891 .n(n)
7892 .k(k)
7893 .cm_stride(19)
7894 .iterations(1)
7895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7896 }
7897 }
7898 }
7899 }
7900
7901 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, a_offset) {
7902 TEST_REQUIRES_ARM_NEON;
7903 for (size_t k = 1; k <= 80; k += 17) {
7904 GemmMicrokernelTester()
7905 .mr(4)
7906 .nr(16)
7907 .kr(2)
7908 .sr(4)
7909 .m(4)
7910 .n(16)
7911 .k(k)
7912 .ks(3)
7913 .a_offset(331)
7914 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7915 }
7916 }
7917
7918 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, zero) {
7919 TEST_REQUIRES_ARM_NEON;
7920 for (size_t k = 1; k <= 80; k += 17) {
7921 for (uint32_t mz = 0; mz < 4; mz++) {
7922 GemmMicrokernelTester()
7923 .mr(4)
7924 .nr(16)
7925 .kr(2)
7926 .sr(4)
7927 .m(4)
7928 .n(16)
7929 .k(k)
7930 .ks(3)
7931 .a_offset(331)
7932 .zero_index(mz)
7933 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7934 }
7935 }
7936 }
7937
7938 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, qmin) {
7939 TEST_REQUIRES_ARM_NEON;
7940 GemmMicrokernelTester()
7941 .mr(4)
7942 .nr(16)
7943 .kr(2)
7944 .sr(4)
7945 .m(4)
7946 .n(16)
7947 .k(16)
7948 .qmin(128)
7949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7950 }
7951
7952 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, qmax) {
7953 TEST_REQUIRES_ARM_NEON;
7954 GemmMicrokernelTester()
7955 .mr(4)
7956 .nr(16)
7957 .kr(2)
7958 .sr(4)
7959 .m(4)
7960 .n(16)
7961 .k(16)
7962 .qmax(128)
7963 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7964 }
7965
7966 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C2S4__NEON_MLAL, strided_cm) {
7967 TEST_REQUIRES_ARM_NEON;
7968 GemmMicrokernelTester()
7969 .mr(4)
7970 .nr(16)
7971 .kr(2)
7972 .sr(4)
7973 .m(4)
7974 .n(16)
7975 .k(16)
7976 .cm_stride(19)
7977 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c2s4__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7978 }
7979#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
7980
7981
7982#if XNN_ARCH_ARM || XNN_ARCH_ARM64
7983 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_eq_8) {
7984 TEST_REQUIRES_ARM_NEON;
7985 GemmMicrokernelTester()
7986 .mr(4)
7987 .nr(8)
7988 .kr(4)
7989 .sr(1)
7990 .m(4)
7991 .n(8)
7992 .k(8)
7993 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
7994 }
7995
7996 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, strided_cn) {
7997 TEST_REQUIRES_ARM_NEON;
7998 GemmMicrokernelTester()
7999 .mr(4)
8000 .nr(8)
8001 .kr(4)
8002 .sr(1)
8003 .m(4)
8004 .n(8)
8005 .k(8)
8006 .cn_stride(11)
8007 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8008 }
8009
8010 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_eq_8_subtile) {
8011 TEST_REQUIRES_ARM_NEON;
8012 for (uint32_t n = 1; n <= 8; n++) {
8013 for (uint32_t m = 1; m <= 4; m++) {
8014 GemmMicrokernelTester()
8015 .mr(4)
8016 .nr(8)
8017 .kr(4)
8018 .sr(1)
8019 .m(m)
8020 .n(n)
8021 .k(8)
8022 .iterations(1)
8023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8024 }
8025 }
8026 }
8027
8028 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_eq_8_subtile_m) {
8029 TEST_REQUIRES_ARM_NEON;
8030 for (uint32_t m = 1; m <= 4; m++) {
8031 GemmMicrokernelTester()
8032 .mr(4)
8033 .nr(8)
8034 .kr(4)
8035 .sr(1)
8036 .m(m)
8037 .n(8)
8038 .k(8)
8039 .iterations(1)
8040 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8041 }
8042 }
8043
8044 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_eq_8_subtile_n) {
8045 TEST_REQUIRES_ARM_NEON;
8046 for (uint32_t n = 1; n <= 8; n++) {
8047 GemmMicrokernelTester()
8048 .mr(4)
8049 .nr(8)
8050 .kr(4)
8051 .sr(1)
8052 .m(4)
8053 .n(n)
8054 .k(8)
8055 .iterations(1)
8056 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8057 }
8058 }
8059
8060 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_lt_8) {
8061 TEST_REQUIRES_ARM_NEON;
8062 for (size_t k = 1; k < 8; k++) {
8063 GemmMicrokernelTester()
8064 .mr(4)
8065 .nr(8)
8066 .kr(4)
8067 .sr(1)
8068 .m(4)
8069 .n(8)
8070 .k(k)
8071 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8072 }
8073 }
8074
8075 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_lt_8_subtile) {
8076 TEST_REQUIRES_ARM_NEON;
8077 for (size_t k = 1; k < 8; k++) {
8078 for (uint32_t n = 1; n <= 8; n++) {
8079 for (uint32_t m = 1; m <= 4; m++) {
8080 GemmMicrokernelTester()
8081 .mr(4)
8082 .nr(8)
8083 .kr(4)
8084 .sr(1)
8085 .m(m)
8086 .n(n)
8087 .k(k)
8088 .iterations(1)
8089 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8090 }
8091 }
8092 }
8093 }
8094
8095 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_gt_8) {
8096 TEST_REQUIRES_ARM_NEON;
8097 for (size_t k = 9; k < 16; k++) {
8098 GemmMicrokernelTester()
8099 .mr(4)
8100 .nr(8)
8101 .kr(4)
8102 .sr(1)
8103 .m(4)
8104 .n(8)
8105 .k(k)
8106 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8107 }
8108 }
8109
8110 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_gt_8_subtile) {
8111 TEST_REQUIRES_ARM_NEON;
8112 for (size_t k = 9; k < 16; k++) {
8113 for (uint32_t n = 1; n <= 8; n++) {
8114 for (uint32_t m = 1; m <= 4; m++) {
8115 GemmMicrokernelTester()
8116 .mr(4)
8117 .nr(8)
8118 .kr(4)
8119 .sr(1)
8120 .m(m)
8121 .n(n)
8122 .k(k)
8123 .iterations(1)
8124 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8125 }
8126 }
8127 }
8128 }
8129
8130 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_div_8) {
8131 TEST_REQUIRES_ARM_NEON;
8132 for (size_t k = 16; k <= 80; k += 8) {
8133 GemmMicrokernelTester()
8134 .mr(4)
8135 .nr(8)
8136 .kr(4)
8137 .sr(1)
8138 .m(4)
8139 .n(8)
8140 .k(k)
8141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8142 }
8143 }
8144
8145 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, k_div_8_subtile) {
8146 TEST_REQUIRES_ARM_NEON;
8147 for (size_t k = 16; k <= 80; k += 8) {
8148 for (uint32_t n = 1; n <= 8; n++) {
8149 for (uint32_t m = 1; m <= 4; m++) {
8150 GemmMicrokernelTester()
8151 .mr(4)
8152 .nr(8)
8153 .kr(4)
8154 .sr(1)
8155 .m(m)
8156 .n(n)
8157 .k(k)
8158 .iterations(1)
8159 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8160 }
8161 }
8162 }
8163 }
8164
8165 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_gt_8) {
8166 TEST_REQUIRES_ARM_NEON;
8167 for (uint32_t n = 9; n < 16; n++) {
8168 for (size_t k = 1; k <= 40; k += 9) {
8169 GemmMicrokernelTester()
8170 .mr(4)
8171 .nr(8)
8172 .kr(4)
8173 .sr(1)
8174 .m(4)
8175 .n(n)
8176 .k(k)
8177 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8178 }
8179 }
8180 }
8181
8182 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_gt_8_strided_cn) {
8183 TEST_REQUIRES_ARM_NEON;
8184 for (uint32_t n = 9; n < 16; n++) {
8185 for (size_t k = 1; k <= 40; k += 9) {
8186 GemmMicrokernelTester()
8187 .mr(4)
8188 .nr(8)
8189 .kr(4)
8190 .sr(1)
8191 .m(4)
8192 .n(n)
8193 .k(k)
8194 .cn_stride(11)
8195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8196 }
8197 }
8198 }
8199
8200 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_gt_8_subtile) {
8201 TEST_REQUIRES_ARM_NEON;
8202 for (uint32_t n = 9; n < 16; n++) {
8203 for (size_t k = 1; k <= 40; k += 9) {
8204 for (uint32_t m = 1; m <= 4; m++) {
8205 GemmMicrokernelTester()
8206 .mr(4)
8207 .nr(8)
8208 .kr(4)
8209 .sr(1)
8210 .m(m)
8211 .n(n)
8212 .k(k)
8213 .iterations(1)
8214 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8215 }
8216 }
8217 }
8218 }
8219
8220 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_div_8) {
8221 TEST_REQUIRES_ARM_NEON;
8222 for (uint32_t n = 16; n <= 24; n += 8) {
8223 for (size_t k = 1; k <= 40; k += 9) {
8224 GemmMicrokernelTester()
8225 .mr(4)
8226 .nr(8)
8227 .kr(4)
8228 .sr(1)
8229 .m(4)
8230 .n(n)
8231 .k(k)
8232 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8233 }
8234 }
8235 }
8236
8237 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_div_8_strided_cn) {
8238 TEST_REQUIRES_ARM_NEON;
8239 for (uint32_t n = 16; n <= 24; n += 8) {
8240 for (size_t k = 1; k <= 40; k += 9) {
8241 GemmMicrokernelTester()
8242 .mr(4)
8243 .nr(8)
8244 .kr(4)
8245 .sr(1)
8246 .m(4)
8247 .n(n)
8248 .k(k)
8249 .cn_stride(11)
8250 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8251 }
8252 }
8253 }
8254
8255 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_div_8_subtile) {
8256 TEST_REQUIRES_ARM_NEON;
8257 for (uint32_t n = 16; n <= 24; n += 8) {
8258 for (size_t k = 1; k <= 40; k += 9) {
8259 for (uint32_t m = 1; m <= 4; m++) {
8260 GemmMicrokernelTester()
8261 .mr(4)
8262 .nr(8)
8263 .kr(4)
8264 .sr(1)
8265 .m(m)
8266 .n(n)
8267 .k(k)
8268 .iterations(1)
8269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8270 }
8271 }
8272 }
8273 }
8274
8275 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, small_kernel) {
8276 TEST_REQUIRES_ARM_NEON;
8277 for (size_t k = 1; k <= 40; k += 9) {
8278 GemmMicrokernelTester()
8279 .mr(4)
8280 .nr(8)
8281 .kr(4)
8282 .sr(1)
8283 .m(4)
8284 .n(8)
8285 .k(k)
8286 .ks(3)
8287 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8288 }
8289 }
8290
8291 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, small_kernel_subtile) {
8292 TEST_REQUIRES_ARM_NEON;
8293 for (size_t k = 1; k <= 40; k += 9) {
8294 for (uint32_t n = 1; n <= 8; n++) {
8295 for (uint32_t m = 1; m <= 4; m++) {
8296 GemmMicrokernelTester()
8297 .mr(4)
8298 .nr(8)
8299 .kr(4)
8300 .sr(1)
8301 .m(m)
8302 .n(n)
8303 .k(k)
8304 .ks(3)
8305 .iterations(1)
8306 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8307 }
8308 }
8309 }
8310 }
8311
8312 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_gt_8_small_kernel) {
8313 TEST_REQUIRES_ARM_NEON;
8314 for (uint32_t n = 9; n < 16; n++) {
8315 for (size_t k = 1; k <= 40; k += 9) {
8316 GemmMicrokernelTester()
8317 .mr(4)
8318 .nr(8)
8319 .kr(4)
8320 .sr(1)
8321 .m(4)
8322 .n(n)
8323 .k(k)
8324 .ks(3)
8325 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8326 }
8327 }
8328 }
8329
8330 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, n_div_8_small_kernel) {
8331 TEST_REQUIRES_ARM_NEON;
8332 for (uint32_t n = 16; n <= 24; n += 8) {
8333 for (size_t k = 1; k <= 40; k += 9) {
8334 GemmMicrokernelTester()
8335 .mr(4)
8336 .nr(8)
8337 .kr(4)
8338 .sr(1)
8339 .m(4)
8340 .n(n)
8341 .k(k)
8342 .ks(3)
8343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8344 }
8345 }
8346 }
8347
8348 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, strided_cm_subtile) {
8349 TEST_REQUIRES_ARM_NEON;
8350 for (size_t k = 1; k <= 40; k += 9) {
8351 for (uint32_t n = 1; n <= 8; n++) {
8352 for (uint32_t m = 1; m <= 4; m++) {
8353 GemmMicrokernelTester()
8354 .mr(4)
8355 .nr(8)
8356 .kr(4)
8357 .sr(1)
8358 .m(m)
8359 .n(n)
8360 .k(k)
8361 .cm_stride(11)
8362 .iterations(1)
8363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8364 }
8365 }
8366 }
8367 }
8368
8369 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, a_offset) {
8370 TEST_REQUIRES_ARM_NEON;
8371 for (size_t k = 1; k <= 40; k += 9) {
8372 GemmMicrokernelTester()
8373 .mr(4)
8374 .nr(8)
8375 .kr(4)
8376 .sr(1)
8377 .m(4)
8378 .n(8)
8379 .k(k)
8380 .ks(3)
8381 .a_offset(163)
8382 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8383 }
8384 }
8385
8386 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, zero) {
8387 TEST_REQUIRES_ARM_NEON;
8388 for (size_t k = 1; k <= 40; k += 9) {
8389 for (uint32_t mz = 0; mz < 4; mz++) {
8390 GemmMicrokernelTester()
8391 .mr(4)
8392 .nr(8)
8393 .kr(4)
8394 .sr(1)
8395 .m(4)
8396 .n(8)
8397 .k(k)
8398 .ks(3)
8399 .a_offset(163)
8400 .zero_index(mz)
8401 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8402 }
8403 }
8404 }
8405
8406 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, qmin) {
8407 TEST_REQUIRES_ARM_NEON;
8408 GemmMicrokernelTester()
8409 .mr(4)
8410 .nr(8)
8411 .kr(4)
8412 .sr(1)
8413 .m(4)
8414 .n(8)
8415 .k(8)
8416 .qmin(128)
8417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8418 }
8419
8420 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, qmax) {
8421 TEST_REQUIRES_ARM_NEON;
8422 GemmMicrokernelTester()
8423 .mr(4)
8424 .nr(8)
8425 .kr(4)
8426 .sr(1)
8427 .m(4)
8428 .n(8)
8429 .k(8)
8430 .qmax(128)
8431 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8432 }
8433
8434 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C4__NEON_MULL_DUP, strided_cm) {
8435 TEST_REQUIRES_ARM_NEON;
8436 GemmMicrokernelTester()
8437 .mr(4)
8438 .nr(8)
8439 .kr(4)
8440 .sr(1)
8441 .m(4)
8442 .n(8)
8443 .k(8)
8444 .cm_stride(11)
8445 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8446 }
8447#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
8448
8449
8450#if XNN_ARCH_ARM || XNN_ARCH_ARM64
8451 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_eq_8) {
8452 TEST_REQUIRES_ARM_NEON;
8453 GemmMicrokernelTester()
8454 .mr(2)
8455 .nr(16)
8456 .kr(4)
8457 .sr(1)
8458 .m(2)
8459 .n(16)
8460 .k(8)
8461 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8462 }
8463
8464 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, strided_cn) {
8465 TEST_REQUIRES_ARM_NEON;
8466 GemmMicrokernelTester()
8467 .mr(2)
8468 .nr(16)
8469 .kr(4)
8470 .sr(1)
8471 .m(2)
8472 .n(16)
8473 .k(8)
8474 .cn_stride(19)
8475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8476 }
8477
8478 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_eq_8_subtile) {
8479 TEST_REQUIRES_ARM_NEON;
8480 for (uint32_t n = 1; n <= 16; n++) {
8481 for (uint32_t m = 1; m <= 2; m++) {
8482 GemmMicrokernelTester()
8483 .mr(2)
8484 .nr(16)
8485 .kr(4)
8486 .sr(1)
8487 .m(m)
8488 .n(n)
8489 .k(8)
8490 .iterations(1)
8491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8492 }
8493 }
8494 }
8495
8496 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_eq_8_subtile_m) {
8497 TEST_REQUIRES_ARM_NEON;
8498 for (uint32_t m = 1; m <= 2; m++) {
8499 GemmMicrokernelTester()
8500 .mr(2)
8501 .nr(16)
8502 .kr(4)
8503 .sr(1)
8504 .m(m)
8505 .n(16)
8506 .k(8)
8507 .iterations(1)
8508 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8509 }
8510 }
8511
8512 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_eq_8_subtile_n) {
8513 TEST_REQUIRES_ARM_NEON;
8514 for (uint32_t n = 1; n <= 16; n++) {
8515 GemmMicrokernelTester()
8516 .mr(2)
8517 .nr(16)
8518 .kr(4)
8519 .sr(1)
8520 .m(2)
8521 .n(n)
8522 .k(8)
8523 .iterations(1)
8524 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8525 }
8526 }
8527
8528 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_lt_8) {
8529 TEST_REQUIRES_ARM_NEON;
8530 for (size_t k = 1; k < 8; k++) {
8531 GemmMicrokernelTester()
8532 .mr(2)
8533 .nr(16)
8534 .kr(4)
8535 .sr(1)
8536 .m(2)
8537 .n(16)
8538 .k(k)
8539 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8540 }
8541 }
8542
8543 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_lt_8_subtile) {
8544 TEST_REQUIRES_ARM_NEON;
8545 for (size_t k = 1; k < 8; k++) {
8546 for (uint32_t n = 1; n <= 16; n++) {
8547 for (uint32_t m = 1; m <= 2; m++) {
8548 GemmMicrokernelTester()
8549 .mr(2)
8550 .nr(16)
8551 .kr(4)
8552 .sr(1)
8553 .m(m)
8554 .n(n)
8555 .k(k)
8556 .iterations(1)
8557 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8558 }
8559 }
8560 }
8561 }
8562
8563 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_gt_8) {
8564 TEST_REQUIRES_ARM_NEON;
8565 for (size_t k = 9; k < 16; k++) {
8566 GemmMicrokernelTester()
8567 .mr(2)
8568 .nr(16)
8569 .kr(4)
8570 .sr(1)
8571 .m(2)
8572 .n(16)
8573 .k(k)
8574 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8575 }
8576 }
8577
8578 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_gt_8_subtile) {
8579 TEST_REQUIRES_ARM_NEON;
8580 for (size_t k = 9; k < 16; k++) {
8581 for (uint32_t n = 1; n <= 16; n++) {
8582 for (uint32_t m = 1; m <= 2; m++) {
8583 GemmMicrokernelTester()
8584 .mr(2)
8585 .nr(16)
8586 .kr(4)
8587 .sr(1)
8588 .m(m)
8589 .n(n)
8590 .k(k)
8591 .iterations(1)
8592 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8593 }
8594 }
8595 }
8596 }
8597
8598 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_div_8) {
8599 TEST_REQUIRES_ARM_NEON;
8600 for (size_t k = 16; k <= 80; k += 8) {
8601 GemmMicrokernelTester()
8602 .mr(2)
8603 .nr(16)
8604 .kr(4)
8605 .sr(1)
8606 .m(2)
8607 .n(16)
8608 .k(k)
8609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8610 }
8611 }
8612
8613 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, k_div_8_subtile) {
8614 TEST_REQUIRES_ARM_NEON;
8615 for (size_t k = 16; k <= 80; k += 8) {
8616 for (uint32_t n = 1; n <= 16; n++) {
8617 for (uint32_t m = 1; m <= 2; m++) {
8618 GemmMicrokernelTester()
8619 .mr(2)
8620 .nr(16)
8621 .kr(4)
8622 .sr(1)
8623 .m(m)
8624 .n(n)
8625 .k(k)
8626 .iterations(1)
8627 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8628 }
8629 }
8630 }
8631 }
8632
8633 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_gt_16) {
8634 TEST_REQUIRES_ARM_NEON;
8635 for (uint32_t n = 17; n < 32; n++) {
8636 for (size_t k = 1; k <= 40; k += 9) {
8637 GemmMicrokernelTester()
8638 .mr(2)
8639 .nr(16)
8640 .kr(4)
8641 .sr(1)
8642 .m(2)
8643 .n(n)
8644 .k(k)
8645 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8646 }
8647 }
8648 }
8649
8650 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_gt_16_strided_cn) {
8651 TEST_REQUIRES_ARM_NEON;
8652 for (uint32_t n = 17; n < 32; n++) {
8653 for (size_t k = 1; k <= 40; k += 9) {
8654 GemmMicrokernelTester()
8655 .mr(2)
8656 .nr(16)
8657 .kr(4)
8658 .sr(1)
8659 .m(2)
8660 .n(n)
8661 .k(k)
8662 .cn_stride(19)
8663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8664 }
8665 }
8666 }
8667
8668 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_gt_16_subtile) {
8669 TEST_REQUIRES_ARM_NEON;
8670 for (uint32_t n = 17; n < 32; n++) {
8671 for (size_t k = 1; k <= 40; k += 9) {
8672 for (uint32_t m = 1; m <= 2; m++) {
8673 GemmMicrokernelTester()
8674 .mr(2)
8675 .nr(16)
8676 .kr(4)
8677 .sr(1)
8678 .m(m)
8679 .n(n)
8680 .k(k)
8681 .iterations(1)
8682 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8683 }
8684 }
8685 }
8686 }
8687
8688 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_div_16) {
8689 TEST_REQUIRES_ARM_NEON;
8690 for (uint32_t n = 32; n <= 48; n += 16) {
8691 for (size_t k = 1; k <= 40; k += 9) {
8692 GemmMicrokernelTester()
8693 .mr(2)
8694 .nr(16)
8695 .kr(4)
8696 .sr(1)
8697 .m(2)
8698 .n(n)
8699 .k(k)
8700 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8701 }
8702 }
8703 }
8704
8705 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_div_16_strided_cn) {
8706 TEST_REQUIRES_ARM_NEON;
8707 for (uint32_t n = 32; n <= 48; n += 16) {
8708 for (size_t k = 1; k <= 40; k += 9) {
8709 GemmMicrokernelTester()
8710 .mr(2)
8711 .nr(16)
8712 .kr(4)
8713 .sr(1)
8714 .m(2)
8715 .n(n)
8716 .k(k)
8717 .cn_stride(19)
8718 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8719 }
8720 }
8721 }
8722
8723 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_div_16_subtile) {
8724 TEST_REQUIRES_ARM_NEON;
8725 for (uint32_t n = 32; n <= 48; n += 16) {
8726 for (size_t k = 1; k <= 40; k += 9) {
8727 for (uint32_t m = 1; m <= 2; m++) {
8728 GemmMicrokernelTester()
8729 .mr(2)
8730 .nr(16)
8731 .kr(4)
8732 .sr(1)
8733 .m(m)
8734 .n(n)
8735 .k(k)
8736 .iterations(1)
8737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8738 }
8739 }
8740 }
8741 }
8742
8743 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, small_kernel) {
8744 TEST_REQUIRES_ARM_NEON;
8745 for (size_t k = 1; k <= 40; k += 9) {
8746 GemmMicrokernelTester()
8747 .mr(2)
8748 .nr(16)
8749 .kr(4)
8750 .sr(1)
8751 .m(2)
8752 .n(16)
8753 .k(k)
8754 .ks(3)
8755 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8756 }
8757 }
8758
8759 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, small_kernel_subtile) {
8760 TEST_REQUIRES_ARM_NEON;
8761 for (size_t k = 1; k <= 40; k += 9) {
8762 for (uint32_t n = 1; n <= 16; n++) {
8763 for (uint32_t m = 1; m <= 2; m++) {
8764 GemmMicrokernelTester()
8765 .mr(2)
8766 .nr(16)
8767 .kr(4)
8768 .sr(1)
8769 .m(m)
8770 .n(n)
8771 .k(k)
8772 .ks(3)
8773 .iterations(1)
8774 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8775 }
8776 }
8777 }
8778 }
8779
8780 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_gt_16_small_kernel) {
8781 TEST_REQUIRES_ARM_NEON;
8782 for (uint32_t n = 17; n < 32; n++) {
8783 for (size_t k = 1; k <= 40; k += 9) {
8784 GemmMicrokernelTester()
8785 .mr(2)
8786 .nr(16)
8787 .kr(4)
8788 .sr(1)
8789 .m(2)
8790 .n(n)
8791 .k(k)
8792 .ks(3)
8793 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8794 }
8795 }
8796 }
8797
8798 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, n_div_16_small_kernel) {
8799 TEST_REQUIRES_ARM_NEON;
8800 for (uint32_t n = 32; n <= 48; n += 16) {
8801 for (size_t k = 1; k <= 40; k += 9) {
8802 GemmMicrokernelTester()
8803 .mr(2)
8804 .nr(16)
8805 .kr(4)
8806 .sr(1)
8807 .m(2)
8808 .n(n)
8809 .k(k)
8810 .ks(3)
8811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8812 }
8813 }
8814 }
8815
8816 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, strided_cm_subtile) {
8817 TEST_REQUIRES_ARM_NEON;
8818 for (size_t k = 1; k <= 40; k += 9) {
8819 for (uint32_t n = 1; n <= 16; n++) {
8820 for (uint32_t m = 1; m <= 2; m++) {
8821 GemmMicrokernelTester()
8822 .mr(2)
8823 .nr(16)
8824 .kr(4)
8825 .sr(1)
8826 .m(m)
8827 .n(n)
8828 .k(k)
8829 .cm_stride(19)
8830 .iterations(1)
8831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8832 }
8833 }
8834 }
8835 }
8836
8837 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, a_offset) {
8838 TEST_REQUIRES_ARM_NEON;
8839 for (size_t k = 1; k <= 40; k += 9) {
8840 GemmMicrokernelTester()
8841 .mr(2)
8842 .nr(16)
8843 .kr(4)
8844 .sr(1)
8845 .m(2)
8846 .n(16)
8847 .k(k)
8848 .ks(3)
8849 .a_offset(83)
8850 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8851 }
8852 }
8853
8854 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, zero) {
8855 TEST_REQUIRES_ARM_NEON;
8856 for (size_t k = 1; k <= 40; k += 9) {
8857 for (uint32_t mz = 0; mz < 2; mz++) {
8858 GemmMicrokernelTester()
8859 .mr(2)
8860 .nr(16)
8861 .kr(4)
8862 .sr(1)
8863 .m(2)
8864 .n(16)
8865 .k(k)
8866 .ks(3)
8867 .a_offset(83)
8868 .zero_index(mz)
8869 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8870 }
8871 }
8872 }
8873
8874 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, qmin) {
8875 TEST_REQUIRES_ARM_NEON;
8876 GemmMicrokernelTester()
8877 .mr(2)
8878 .nr(16)
8879 .kr(4)
8880 .sr(1)
8881 .m(2)
8882 .n(16)
8883 .k(8)
8884 .qmin(128)
8885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8886 }
8887
8888 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, qmax) {
8889 TEST_REQUIRES_ARM_NEON;
8890 GemmMicrokernelTester()
8891 .mr(2)
8892 .nr(16)
8893 .kr(4)
8894 .sr(1)
8895 .m(2)
8896 .n(16)
8897 .k(8)
8898 .qmax(128)
8899 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8900 }
8901
8902 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C4__NEON_MULL_DUP, strided_cm) {
8903 TEST_REQUIRES_ARM_NEON;
8904 GemmMicrokernelTester()
8905 .mr(2)
8906 .nr(16)
8907 .kr(4)
8908 .sr(1)
8909 .m(2)
8910 .n(16)
8911 .k(8)
8912 .cm_stride(19)
8913 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c4__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8914 }
8915#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
8916
8917
8918#if XNN_ARCH_ARM || XNN_ARCH_ARM64
8919 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_eq_16) {
8920 TEST_REQUIRES_ARM_NEON;
8921 GemmMicrokernelTester()
8922 .mr(1)
8923 .nr(8)
8924 .kr(4)
8925 .sr(1)
8926 .m(1)
8927 .n(8)
8928 .k(16)
8929 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8930 }
8931
8932 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, strided_cn) {
8933 TEST_REQUIRES_ARM_NEON;
8934 GemmMicrokernelTester()
8935 .mr(1)
8936 .nr(8)
8937 .kr(4)
8938 .sr(1)
8939 .m(1)
8940 .n(8)
8941 .k(16)
8942 .cn_stride(11)
8943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8944 }
8945
8946 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_eq_16_subtile) {
8947 TEST_REQUIRES_ARM_NEON;
8948 for (uint32_t n = 1; n <= 8; n++) {
8949 for (uint32_t m = 1; m <= 1; m++) {
8950 GemmMicrokernelTester()
8951 .mr(1)
8952 .nr(8)
8953 .kr(4)
8954 .sr(1)
8955 .m(m)
8956 .n(n)
8957 .k(16)
8958 .iterations(1)
8959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8960 }
8961 }
8962 }
8963
8964 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_eq_16_subtile_m) {
8965 TEST_REQUIRES_ARM_NEON;
8966 for (uint32_t m = 1; m <= 1; m++) {
8967 GemmMicrokernelTester()
8968 .mr(1)
8969 .nr(8)
8970 .kr(4)
8971 .sr(1)
8972 .m(m)
8973 .n(8)
8974 .k(16)
8975 .iterations(1)
8976 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8977 }
8978 }
8979
8980 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_eq_16_subtile_n) {
8981 TEST_REQUIRES_ARM_NEON;
8982 for (uint32_t n = 1; n <= 8; n++) {
8983 GemmMicrokernelTester()
8984 .mr(1)
8985 .nr(8)
8986 .kr(4)
8987 .sr(1)
8988 .m(1)
8989 .n(n)
8990 .k(16)
8991 .iterations(1)
8992 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
8993 }
8994 }
8995
8996 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_lt_16) {
8997 TEST_REQUIRES_ARM_NEON;
8998 for (size_t k = 1; k < 16; k++) {
8999 GemmMicrokernelTester()
9000 .mr(1)
9001 .nr(8)
9002 .kr(4)
9003 .sr(1)
9004 .m(1)
9005 .n(8)
9006 .k(k)
9007 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9008 }
9009 }
9010
9011 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_lt_16_subtile) {
9012 TEST_REQUIRES_ARM_NEON;
9013 for (size_t k = 1; k < 16; k++) {
9014 for (uint32_t n = 1; n <= 8; n++) {
9015 for (uint32_t m = 1; m <= 1; m++) {
9016 GemmMicrokernelTester()
9017 .mr(1)
9018 .nr(8)
9019 .kr(4)
9020 .sr(1)
9021 .m(m)
9022 .n(n)
9023 .k(k)
9024 .iterations(1)
9025 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9026 }
9027 }
9028 }
9029 }
9030
9031 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_gt_16) {
9032 TEST_REQUIRES_ARM_NEON;
9033 for (size_t k = 17; k < 32; k++) {
9034 GemmMicrokernelTester()
9035 .mr(1)
9036 .nr(8)
9037 .kr(4)
9038 .sr(1)
9039 .m(1)
9040 .n(8)
9041 .k(k)
9042 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9043 }
9044 }
9045
9046 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_gt_16_subtile) {
9047 TEST_REQUIRES_ARM_NEON;
9048 for (size_t k = 17; k < 32; k++) {
9049 for (uint32_t n = 1; n <= 8; n++) {
9050 for (uint32_t m = 1; m <= 1; m++) {
9051 GemmMicrokernelTester()
9052 .mr(1)
9053 .nr(8)
9054 .kr(4)
9055 .sr(1)
9056 .m(m)
9057 .n(n)
9058 .k(k)
9059 .iterations(1)
9060 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9061 }
9062 }
9063 }
9064 }
9065
9066 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_div_16) {
9067 TEST_REQUIRES_ARM_NEON;
9068 for (size_t k = 32; k <= 160; k += 16) {
9069 GemmMicrokernelTester()
9070 .mr(1)
9071 .nr(8)
9072 .kr(4)
9073 .sr(1)
9074 .m(1)
9075 .n(8)
9076 .k(k)
9077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9078 }
9079 }
9080
9081 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, k_div_16_subtile) {
9082 TEST_REQUIRES_ARM_NEON;
9083 for (size_t k = 32; k <= 160; k += 16) {
9084 for (uint32_t n = 1; n <= 8; n++) {
9085 for (uint32_t m = 1; m <= 1; m++) {
9086 GemmMicrokernelTester()
9087 .mr(1)
9088 .nr(8)
9089 .kr(4)
9090 .sr(1)
9091 .m(m)
9092 .n(n)
9093 .k(k)
9094 .iterations(1)
9095 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9096 }
9097 }
9098 }
9099 }
9100
9101 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_gt_8) {
9102 TEST_REQUIRES_ARM_NEON;
9103 for (uint32_t n = 9; n < 16; n++) {
9104 for (size_t k = 1; k <= 80; k += 17) {
9105 GemmMicrokernelTester()
9106 .mr(1)
9107 .nr(8)
9108 .kr(4)
9109 .sr(1)
9110 .m(1)
9111 .n(n)
9112 .k(k)
9113 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9114 }
9115 }
9116 }
9117
9118 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_gt_8_strided_cn) {
9119 TEST_REQUIRES_ARM_NEON;
9120 for (uint32_t n = 9; n < 16; n++) {
9121 for (size_t k = 1; k <= 80; k += 17) {
9122 GemmMicrokernelTester()
9123 .mr(1)
9124 .nr(8)
9125 .kr(4)
9126 .sr(1)
9127 .m(1)
9128 .n(n)
9129 .k(k)
9130 .cn_stride(11)
9131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9132 }
9133 }
9134 }
9135
9136 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_gt_8_subtile) {
9137 TEST_REQUIRES_ARM_NEON;
9138 for (uint32_t n = 9; n < 16; n++) {
9139 for (size_t k = 1; k <= 80; k += 17) {
9140 for (uint32_t m = 1; m <= 1; m++) {
9141 GemmMicrokernelTester()
9142 .mr(1)
9143 .nr(8)
9144 .kr(4)
9145 .sr(1)
9146 .m(m)
9147 .n(n)
9148 .k(k)
9149 .iterations(1)
9150 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9151 }
9152 }
9153 }
9154 }
9155
9156 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_div_8) {
9157 TEST_REQUIRES_ARM_NEON;
9158 for (uint32_t n = 16; n <= 24; n += 8) {
9159 for (size_t k = 1; k <= 80; k += 17) {
9160 GemmMicrokernelTester()
9161 .mr(1)
9162 .nr(8)
9163 .kr(4)
9164 .sr(1)
9165 .m(1)
9166 .n(n)
9167 .k(k)
9168 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9169 }
9170 }
9171 }
9172
9173 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_div_8_strided_cn) {
9174 TEST_REQUIRES_ARM_NEON;
9175 for (uint32_t n = 16; n <= 24; n += 8) {
9176 for (size_t k = 1; k <= 80; k += 17) {
9177 GemmMicrokernelTester()
9178 .mr(1)
9179 .nr(8)
9180 .kr(4)
9181 .sr(1)
9182 .m(1)
9183 .n(n)
9184 .k(k)
9185 .cn_stride(11)
9186 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9187 }
9188 }
9189 }
9190
9191 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_div_8_subtile) {
9192 TEST_REQUIRES_ARM_NEON;
9193 for (uint32_t n = 16; n <= 24; n += 8) {
9194 for (size_t k = 1; k <= 80; k += 17) {
9195 for (uint32_t m = 1; m <= 1; m++) {
9196 GemmMicrokernelTester()
9197 .mr(1)
9198 .nr(8)
9199 .kr(4)
9200 .sr(1)
9201 .m(m)
9202 .n(n)
9203 .k(k)
9204 .iterations(1)
9205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9206 }
9207 }
9208 }
9209 }
9210
9211 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, small_kernel) {
9212 TEST_REQUIRES_ARM_NEON;
9213 for (size_t k = 1; k <= 80; k += 17) {
9214 GemmMicrokernelTester()
9215 .mr(1)
9216 .nr(8)
9217 .kr(4)
9218 .sr(1)
9219 .m(1)
9220 .n(8)
9221 .k(k)
9222 .ks(3)
9223 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9224 }
9225 }
9226
9227 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, small_kernel_subtile) {
9228 TEST_REQUIRES_ARM_NEON;
9229 for (size_t k = 1; k <= 80; k += 17) {
9230 for (uint32_t n = 1; n <= 8; n++) {
9231 for (uint32_t m = 1; m <= 1; m++) {
9232 GemmMicrokernelTester()
9233 .mr(1)
9234 .nr(8)
9235 .kr(4)
9236 .sr(1)
9237 .m(m)
9238 .n(n)
9239 .k(k)
9240 .ks(3)
9241 .iterations(1)
9242 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9243 }
9244 }
9245 }
9246 }
9247
9248 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_gt_8_small_kernel) {
9249 TEST_REQUIRES_ARM_NEON;
9250 for (uint32_t n = 9; n < 16; n++) {
9251 for (size_t k = 1; k <= 80; k += 17) {
9252 GemmMicrokernelTester()
9253 .mr(1)
9254 .nr(8)
9255 .kr(4)
9256 .sr(1)
9257 .m(1)
9258 .n(n)
9259 .k(k)
9260 .ks(3)
9261 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9262 }
9263 }
9264 }
9265
9266 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, n_div_8_small_kernel) {
9267 TEST_REQUIRES_ARM_NEON;
9268 for (uint32_t n = 16; n <= 24; n += 8) {
9269 for (size_t k = 1; k <= 80; k += 17) {
9270 GemmMicrokernelTester()
9271 .mr(1)
9272 .nr(8)
9273 .kr(4)
9274 .sr(1)
9275 .m(1)
9276 .n(n)
9277 .k(k)
9278 .ks(3)
9279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9280 }
9281 }
9282 }
9283
9284 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, strided_cm_subtile) {
9285 TEST_REQUIRES_ARM_NEON;
9286 for (size_t k = 1; k <= 80; k += 17) {
9287 for (uint32_t n = 1; n <= 8; n++) {
9288 for (uint32_t m = 1; m <= 1; m++) {
9289 GemmMicrokernelTester()
9290 .mr(1)
9291 .nr(8)
9292 .kr(4)
9293 .sr(1)
9294 .m(m)
9295 .n(n)
9296 .k(k)
9297 .cm_stride(11)
9298 .iterations(1)
9299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9300 }
9301 }
9302 }
9303 }
9304
9305 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, a_offset) {
9306 TEST_REQUIRES_ARM_NEON;
9307 for (size_t k = 1; k <= 80; k += 17) {
9308 GemmMicrokernelTester()
9309 .mr(1)
9310 .nr(8)
9311 .kr(4)
9312 .sr(1)
9313 .m(1)
9314 .n(8)
9315 .k(k)
9316 .ks(3)
9317 .a_offset(83)
9318 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9319 }
9320 }
9321
9322 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, zero) {
9323 TEST_REQUIRES_ARM_NEON;
9324 for (size_t k = 1; k <= 80; k += 17) {
9325 for (uint32_t mz = 0; mz < 1; mz++) {
9326 GemmMicrokernelTester()
9327 .mr(1)
9328 .nr(8)
9329 .kr(4)
9330 .sr(1)
9331 .m(1)
9332 .n(8)
9333 .k(k)
9334 .ks(3)
9335 .a_offset(83)
9336 .zero_index(mz)
9337 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9338 }
9339 }
9340 }
9341
9342 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, qmin) {
9343 TEST_REQUIRES_ARM_NEON;
9344 GemmMicrokernelTester()
9345 .mr(1)
9346 .nr(8)
9347 .kr(4)
9348 .sr(1)
9349 .m(1)
9350 .n(8)
9351 .k(16)
9352 .qmin(128)
9353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9354 }
9355
9356 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, qmax) {
9357 TEST_REQUIRES_ARM_NEON;
9358 GemmMicrokernelTester()
9359 .mr(1)
9360 .nr(8)
9361 .kr(4)
9362 .sr(1)
9363 .m(1)
9364 .n(8)
9365 .k(16)
9366 .qmax(128)
9367 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9368 }
9369
9370 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_DUP, strided_cm) {
9371 TEST_REQUIRES_ARM_NEON;
9372 GemmMicrokernelTester()
9373 .mr(1)
9374 .nr(8)
9375 .kr(4)
9376 .sr(1)
9377 .m(1)
9378 .n(8)
9379 .k(16)
9380 .cm_stride(11)
9381 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9382 }
9383#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
9384
9385
9386#if XNN_ARCH_ARM || XNN_ARCH_ARM64
9387 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16) {
9388 TEST_REQUIRES_ARM_NEON;
9389 GemmMicrokernelTester()
9390 .mr(3)
9391 .nr(8)
9392 .kr(4)
9393 .sr(1)
9394 .m(3)
9395 .n(8)
9396 .k(16)
9397 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9398 }
9399
9400 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, strided_cn) {
9401 TEST_REQUIRES_ARM_NEON;
9402 GemmMicrokernelTester()
9403 .mr(3)
9404 .nr(8)
9405 .kr(4)
9406 .sr(1)
9407 .m(3)
9408 .n(8)
9409 .k(16)
9410 .cn_stride(11)
9411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9412 }
9413
9414 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16_subtile) {
9415 TEST_REQUIRES_ARM_NEON;
9416 for (uint32_t n = 1; n <= 8; n++) {
9417 for (uint32_t m = 1; m <= 3; m++) {
9418 GemmMicrokernelTester()
9419 .mr(3)
9420 .nr(8)
9421 .kr(4)
9422 .sr(1)
9423 .m(m)
9424 .n(n)
9425 .k(16)
9426 .iterations(1)
9427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9428 }
9429 }
9430 }
9431
9432 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16_subtile_m) {
9433 TEST_REQUIRES_ARM_NEON;
9434 for (uint32_t m = 1; m <= 3; m++) {
9435 GemmMicrokernelTester()
9436 .mr(3)
9437 .nr(8)
9438 .kr(4)
9439 .sr(1)
9440 .m(m)
9441 .n(8)
9442 .k(16)
9443 .iterations(1)
9444 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9445 }
9446 }
9447
9448 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_eq_16_subtile_n) {
9449 TEST_REQUIRES_ARM_NEON;
9450 for (uint32_t n = 1; n <= 8; n++) {
9451 GemmMicrokernelTester()
9452 .mr(3)
9453 .nr(8)
9454 .kr(4)
9455 .sr(1)
9456 .m(3)
9457 .n(n)
9458 .k(16)
9459 .iterations(1)
9460 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9461 }
9462 }
9463
9464 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_lt_16) {
9465 TEST_REQUIRES_ARM_NEON;
9466 for (size_t k = 1; k < 16; k++) {
9467 GemmMicrokernelTester()
9468 .mr(3)
9469 .nr(8)
9470 .kr(4)
9471 .sr(1)
9472 .m(3)
9473 .n(8)
9474 .k(k)
9475 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9476 }
9477 }
9478
9479 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_lt_16_subtile) {
9480 TEST_REQUIRES_ARM_NEON;
9481 for (size_t k = 1; k < 16; k++) {
9482 for (uint32_t n = 1; n <= 8; n++) {
9483 for (uint32_t m = 1; m <= 3; m++) {
9484 GemmMicrokernelTester()
9485 .mr(3)
9486 .nr(8)
9487 .kr(4)
9488 .sr(1)
9489 .m(m)
9490 .n(n)
9491 .k(k)
9492 .iterations(1)
9493 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9494 }
9495 }
9496 }
9497 }
9498
9499 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_gt_16) {
9500 TEST_REQUIRES_ARM_NEON;
9501 for (size_t k = 17; k < 32; k++) {
9502 GemmMicrokernelTester()
9503 .mr(3)
9504 .nr(8)
9505 .kr(4)
9506 .sr(1)
9507 .m(3)
9508 .n(8)
9509 .k(k)
9510 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9511 }
9512 }
9513
9514 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_gt_16_subtile) {
9515 TEST_REQUIRES_ARM_NEON;
9516 for (size_t k = 17; k < 32; k++) {
9517 for (uint32_t n = 1; n <= 8; n++) {
9518 for (uint32_t m = 1; m <= 3; m++) {
9519 GemmMicrokernelTester()
9520 .mr(3)
9521 .nr(8)
9522 .kr(4)
9523 .sr(1)
9524 .m(m)
9525 .n(n)
9526 .k(k)
9527 .iterations(1)
9528 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9529 }
9530 }
9531 }
9532 }
9533
9534 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_div_16) {
9535 TEST_REQUIRES_ARM_NEON;
9536 for (size_t k = 32; k <= 160; k += 16) {
9537 GemmMicrokernelTester()
9538 .mr(3)
9539 .nr(8)
9540 .kr(4)
9541 .sr(1)
9542 .m(3)
9543 .n(8)
9544 .k(k)
9545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9546 }
9547 }
9548
9549 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, k_div_16_subtile) {
9550 TEST_REQUIRES_ARM_NEON;
9551 for (size_t k = 32; k <= 160; k += 16) {
9552 for (uint32_t n = 1; n <= 8; n++) {
9553 for (uint32_t m = 1; m <= 3; m++) {
9554 GemmMicrokernelTester()
9555 .mr(3)
9556 .nr(8)
9557 .kr(4)
9558 .sr(1)
9559 .m(m)
9560 .n(n)
9561 .k(k)
9562 .iterations(1)
9563 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9564 }
9565 }
9566 }
9567 }
9568
9569 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8) {
9570 TEST_REQUIRES_ARM_NEON;
9571 for (uint32_t n = 9; n < 16; n++) {
9572 for (size_t k = 1; k <= 80; k += 17) {
9573 GemmMicrokernelTester()
9574 .mr(3)
9575 .nr(8)
9576 .kr(4)
9577 .sr(1)
9578 .m(3)
9579 .n(n)
9580 .k(k)
9581 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9582 }
9583 }
9584 }
9585
9586 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8_strided_cn) {
9587 TEST_REQUIRES_ARM_NEON;
9588 for (uint32_t n = 9; n < 16; n++) {
9589 for (size_t k = 1; k <= 80; k += 17) {
9590 GemmMicrokernelTester()
9591 .mr(3)
9592 .nr(8)
9593 .kr(4)
9594 .sr(1)
9595 .m(3)
9596 .n(n)
9597 .k(k)
9598 .cn_stride(11)
9599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9600 }
9601 }
9602 }
9603
9604 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8_subtile) {
9605 TEST_REQUIRES_ARM_NEON;
9606 for (uint32_t n = 9; n < 16; n++) {
9607 for (size_t k = 1; k <= 80; k += 17) {
9608 for (uint32_t m = 1; m <= 3; m++) {
9609 GemmMicrokernelTester()
9610 .mr(3)
9611 .nr(8)
9612 .kr(4)
9613 .sr(1)
9614 .m(m)
9615 .n(n)
9616 .k(k)
9617 .iterations(1)
9618 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9619 }
9620 }
9621 }
9622 }
9623
9624 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8) {
9625 TEST_REQUIRES_ARM_NEON;
9626 for (uint32_t n = 16; n <= 24; n += 8) {
9627 for (size_t k = 1; k <= 80; k += 17) {
9628 GemmMicrokernelTester()
9629 .mr(3)
9630 .nr(8)
9631 .kr(4)
9632 .sr(1)
9633 .m(3)
9634 .n(n)
9635 .k(k)
9636 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9637 }
9638 }
9639 }
9640
9641 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8_strided_cn) {
9642 TEST_REQUIRES_ARM_NEON;
9643 for (uint32_t n = 16; n <= 24; n += 8) {
9644 for (size_t k = 1; k <= 80; k += 17) {
9645 GemmMicrokernelTester()
9646 .mr(3)
9647 .nr(8)
9648 .kr(4)
9649 .sr(1)
9650 .m(3)
9651 .n(n)
9652 .k(k)
9653 .cn_stride(11)
9654 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9655 }
9656 }
9657 }
9658
9659 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8_subtile) {
9660 TEST_REQUIRES_ARM_NEON;
9661 for (uint32_t n = 16; n <= 24; n += 8) {
9662 for (size_t k = 1; k <= 80; k += 17) {
9663 for (uint32_t m = 1; m <= 3; m++) {
9664 GemmMicrokernelTester()
9665 .mr(3)
9666 .nr(8)
9667 .kr(4)
9668 .sr(1)
9669 .m(m)
9670 .n(n)
9671 .k(k)
9672 .iterations(1)
9673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9674 }
9675 }
9676 }
9677 }
9678
9679 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, small_kernel) {
9680 TEST_REQUIRES_ARM_NEON;
9681 for (size_t k = 1; k <= 80; k += 17) {
9682 GemmMicrokernelTester()
9683 .mr(3)
9684 .nr(8)
9685 .kr(4)
9686 .sr(1)
9687 .m(3)
9688 .n(8)
9689 .k(k)
9690 .ks(3)
9691 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9692 }
9693 }
9694
9695 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, small_kernel_subtile) {
9696 TEST_REQUIRES_ARM_NEON;
9697 for (size_t k = 1; k <= 80; k += 17) {
9698 for (uint32_t n = 1; n <= 8; n++) {
9699 for (uint32_t m = 1; m <= 3; m++) {
9700 GemmMicrokernelTester()
9701 .mr(3)
9702 .nr(8)
9703 .kr(4)
9704 .sr(1)
9705 .m(m)
9706 .n(n)
9707 .k(k)
9708 .ks(3)
9709 .iterations(1)
9710 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9711 }
9712 }
9713 }
9714 }
9715
9716 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_gt_8_small_kernel) {
9717 TEST_REQUIRES_ARM_NEON;
9718 for (uint32_t n = 9; n < 16; n++) {
9719 for (size_t k = 1; k <= 80; k += 17) {
9720 GemmMicrokernelTester()
9721 .mr(3)
9722 .nr(8)
9723 .kr(4)
9724 .sr(1)
9725 .m(3)
9726 .n(n)
9727 .k(k)
9728 .ks(3)
9729 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9730 }
9731 }
9732 }
9733
9734 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, n_div_8_small_kernel) {
9735 TEST_REQUIRES_ARM_NEON;
9736 for (uint32_t n = 16; n <= 24; n += 8) {
9737 for (size_t k = 1; k <= 80; k += 17) {
9738 GemmMicrokernelTester()
9739 .mr(3)
9740 .nr(8)
9741 .kr(4)
9742 .sr(1)
9743 .m(3)
9744 .n(n)
9745 .k(k)
9746 .ks(3)
9747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9748 }
9749 }
9750 }
9751
9752 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, strided_cm_subtile) {
9753 TEST_REQUIRES_ARM_NEON;
9754 for (size_t k = 1; k <= 80; k += 17) {
9755 for (uint32_t n = 1; n <= 8; n++) {
9756 for (uint32_t m = 1; m <= 3; m++) {
9757 GemmMicrokernelTester()
9758 .mr(3)
9759 .nr(8)
9760 .kr(4)
9761 .sr(1)
9762 .m(m)
9763 .n(n)
9764 .k(k)
9765 .cm_stride(11)
9766 .iterations(1)
9767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9768 }
9769 }
9770 }
9771 }
9772
9773 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, a_offset) {
9774 TEST_REQUIRES_ARM_NEON;
9775 for (size_t k = 1; k <= 80; k += 17) {
9776 GemmMicrokernelTester()
9777 .mr(3)
9778 .nr(8)
9779 .kr(4)
9780 .sr(1)
9781 .m(3)
9782 .n(8)
9783 .k(k)
9784 .ks(3)
9785 .a_offset(251)
9786 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9787 }
9788 }
9789
9790 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, zero) {
9791 TEST_REQUIRES_ARM_NEON;
9792 for (size_t k = 1; k <= 80; k += 17) {
9793 for (uint32_t mz = 0; mz < 3; mz++) {
9794 GemmMicrokernelTester()
9795 .mr(3)
9796 .nr(8)
9797 .kr(4)
9798 .sr(1)
9799 .m(3)
9800 .n(8)
9801 .k(k)
9802 .ks(3)
9803 .a_offset(251)
9804 .zero_index(mz)
9805 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9806 }
9807 }
9808 }
9809
9810 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, qmin) {
9811 TEST_REQUIRES_ARM_NEON;
9812 GemmMicrokernelTester()
9813 .mr(3)
9814 .nr(8)
9815 .kr(4)
9816 .sr(1)
9817 .m(3)
9818 .n(8)
9819 .k(16)
9820 .qmin(128)
9821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9822 }
9823
9824 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, qmax) {
9825 TEST_REQUIRES_ARM_NEON;
9826 GemmMicrokernelTester()
9827 .mr(3)
9828 .nr(8)
9829 .kr(4)
9830 .sr(1)
9831 .m(3)
9832 .n(8)
9833 .k(16)
9834 .qmax(128)
9835 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9836 }
9837
9838 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C4__NEON_MLAL_DUP, strided_cm) {
9839 TEST_REQUIRES_ARM_NEON;
9840 GemmMicrokernelTester()
9841 .mr(3)
9842 .nr(8)
9843 .kr(4)
9844 .sr(1)
9845 .m(3)
9846 .n(8)
9847 .k(16)
9848 .cm_stride(11)
9849 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9850 }
9851#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
9852
9853
9854#if XNN_ARCH_ARM || XNN_ARCH_ARM64
9855 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_eq_16) {
9856 TEST_REQUIRES_ARM_NEON;
9857 GemmMicrokernelTester()
9858 .mr(1)
9859 .nr(16)
9860 .kr(4)
9861 .sr(1)
9862 .m(1)
9863 .n(16)
9864 .k(16)
9865 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9866 }
9867
9868 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, strided_cn) {
9869 TEST_REQUIRES_ARM_NEON;
9870 GemmMicrokernelTester()
9871 .mr(1)
9872 .nr(16)
9873 .kr(4)
9874 .sr(1)
9875 .m(1)
9876 .n(16)
9877 .k(16)
9878 .cn_stride(19)
9879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9880 }
9881
9882 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_eq_16_subtile) {
9883 TEST_REQUIRES_ARM_NEON;
9884 for (uint32_t n = 1; n <= 16; n++) {
9885 for (uint32_t m = 1; m <= 1; m++) {
9886 GemmMicrokernelTester()
9887 .mr(1)
9888 .nr(16)
9889 .kr(4)
9890 .sr(1)
9891 .m(m)
9892 .n(n)
9893 .k(16)
9894 .iterations(1)
9895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9896 }
9897 }
9898 }
9899
9900 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_eq_16_subtile_m) {
9901 TEST_REQUIRES_ARM_NEON;
9902 for (uint32_t m = 1; m <= 1; m++) {
9903 GemmMicrokernelTester()
9904 .mr(1)
9905 .nr(16)
9906 .kr(4)
9907 .sr(1)
9908 .m(m)
9909 .n(16)
9910 .k(16)
9911 .iterations(1)
9912 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9913 }
9914 }
9915
9916 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_eq_16_subtile_n) {
9917 TEST_REQUIRES_ARM_NEON;
9918 for (uint32_t n = 1; n <= 16; n++) {
9919 GemmMicrokernelTester()
9920 .mr(1)
9921 .nr(16)
9922 .kr(4)
9923 .sr(1)
9924 .m(1)
9925 .n(n)
9926 .k(16)
9927 .iterations(1)
9928 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9929 }
9930 }
9931
9932 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_lt_16) {
9933 TEST_REQUIRES_ARM_NEON;
9934 for (size_t k = 1; k < 16; k++) {
9935 GemmMicrokernelTester()
9936 .mr(1)
9937 .nr(16)
9938 .kr(4)
9939 .sr(1)
9940 .m(1)
9941 .n(16)
9942 .k(k)
9943 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9944 }
9945 }
9946
9947 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_lt_16_subtile) {
9948 TEST_REQUIRES_ARM_NEON;
9949 for (size_t k = 1; k < 16; k++) {
9950 for (uint32_t n = 1; n <= 16; n++) {
9951 for (uint32_t m = 1; m <= 1; m++) {
9952 GemmMicrokernelTester()
9953 .mr(1)
9954 .nr(16)
9955 .kr(4)
9956 .sr(1)
9957 .m(m)
9958 .n(n)
9959 .k(k)
9960 .iterations(1)
9961 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9962 }
9963 }
9964 }
9965 }
9966
9967 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_gt_16) {
9968 TEST_REQUIRES_ARM_NEON;
9969 for (size_t k = 17; k < 32; k++) {
9970 GemmMicrokernelTester()
9971 .mr(1)
9972 .nr(16)
9973 .kr(4)
9974 .sr(1)
9975 .m(1)
9976 .n(16)
9977 .k(k)
9978 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9979 }
9980 }
9981
9982 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_gt_16_subtile) {
9983 TEST_REQUIRES_ARM_NEON;
9984 for (size_t k = 17; k < 32; k++) {
9985 for (uint32_t n = 1; n <= 16; n++) {
9986 for (uint32_t m = 1; m <= 1; m++) {
9987 GemmMicrokernelTester()
9988 .mr(1)
9989 .nr(16)
9990 .kr(4)
9991 .sr(1)
9992 .m(m)
9993 .n(n)
9994 .k(k)
9995 .iterations(1)
9996 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
9997 }
9998 }
9999 }
10000 }
10001
10002 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_div_16) {
10003 TEST_REQUIRES_ARM_NEON;
10004 for (size_t k = 32; k <= 160; k += 16) {
10005 GemmMicrokernelTester()
10006 .mr(1)
10007 .nr(16)
10008 .kr(4)
10009 .sr(1)
10010 .m(1)
10011 .n(16)
10012 .k(k)
10013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10014 }
10015 }
10016
10017 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, k_div_16_subtile) {
10018 TEST_REQUIRES_ARM_NEON;
10019 for (size_t k = 32; k <= 160; k += 16) {
10020 for (uint32_t n = 1; n <= 16; n++) {
10021 for (uint32_t m = 1; m <= 1; m++) {
10022 GemmMicrokernelTester()
10023 .mr(1)
10024 .nr(16)
10025 .kr(4)
10026 .sr(1)
10027 .m(m)
10028 .n(n)
10029 .k(k)
10030 .iterations(1)
10031 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10032 }
10033 }
10034 }
10035 }
10036
10037 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_gt_16) {
10038 TEST_REQUIRES_ARM_NEON;
10039 for (uint32_t n = 17; n < 32; n++) {
10040 for (size_t k = 1; k <= 80; k += 17) {
10041 GemmMicrokernelTester()
10042 .mr(1)
10043 .nr(16)
10044 .kr(4)
10045 .sr(1)
10046 .m(1)
10047 .n(n)
10048 .k(k)
10049 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10050 }
10051 }
10052 }
10053
10054 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_gt_16_strided_cn) {
10055 TEST_REQUIRES_ARM_NEON;
10056 for (uint32_t n = 17; n < 32; n++) {
10057 for (size_t k = 1; k <= 80; k += 17) {
10058 GemmMicrokernelTester()
10059 .mr(1)
10060 .nr(16)
10061 .kr(4)
10062 .sr(1)
10063 .m(1)
10064 .n(n)
10065 .k(k)
10066 .cn_stride(19)
10067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10068 }
10069 }
10070 }
10071
10072 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_gt_16_subtile) {
10073 TEST_REQUIRES_ARM_NEON;
10074 for (uint32_t n = 17; n < 32; n++) {
10075 for (size_t k = 1; k <= 80; k += 17) {
10076 for (uint32_t m = 1; m <= 1; m++) {
10077 GemmMicrokernelTester()
10078 .mr(1)
10079 .nr(16)
10080 .kr(4)
10081 .sr(1)
10082 .m(m)
10083 .n(n)
10084 .k(k)
10085 .iterations(1)
10086 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10087 }
10088 }
10089 }
10090 }
10091
10092 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_div_16) {
10093 TEST_REQUIRES_ARM_NEON;
10094 for (uint32_t n = 32; n <= 48; n += 16) {
10095 for (size_t k = 1; k <= 80; k += 17) {
10096 GemmMicrokernelTester()
10097 .mr(1)
10098 .nr(16)
10099 .kr(4)
10100 .sr(1)
10101 .m(1)
10102 .n(n)
10103 .k(k)
10104 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10105 }
10106 }
10107 }
10108
10109 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_div_16_strided_cn) {
10110 TEST_REQUIRES_ARM_NEON;
10111 for (uint32_t n = 32; n <= 48; n += 16) {
10112 for (size_t k = 1; k <= 80; k += 17) {
10113 GemmMicrokernelTester()
10114 .mr(1)
10115 .nr(16)
10116 .kr(4)
10117 .sr(1)
10118 .m(1)
10119 .n(n)
10120 .k(k)
10121 .cn_stride(19)
10122 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10123 }
10124 }
10125 }
10126
10127 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_div_16_subtile) {
10128 TEST_REQUIRES_ARM_NEON;
10129 for (uint32_t n = 32; n <= 48; n += 16) {
10130 for (size_t k = 1; k <= 80; k += 17) {
10131 for (uint32_t m = 1; m <= 1; m++) {
10132 GemmMicrokernelTester()
10133 .mr(1)
10134 .nr(16)
10135 .kr(4)
10136 .sr(1)
10137 .m(m)
10138 .n(n)
10139 .k(k)
10140 .iterations(1)
10141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10142 }
10143 }
10144 }
10145 }
10146
10147 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, small_kernel) {
10148 TEST_REQUIRES_ARM_NEON;
10149 for (size_t k = 1; k <= 80; k += 17) {
10150 GemmMicrokernelTester()
10151 .mr(1)
10152 .nr(16)
10153 .kr(4)
10154 .sr(1)
10155 .m(1)
10156 .n(16)
10157 .k(k)
10158 .ks(3)
10159 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10160 }
10161 }
10162
10163 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, small_kernel_subtile) {
10164 TEST_REQUIRES_ARM_NEON;
10165 for (size_t k = 1; k <= 80; k += 17) {
10166 for (uint32_t n = 1; n <= 16; n++) {
10167 for (uint32_t m = 1; m <= 1; m++) {
10168 GemmMicrokernelTester()
10169 .mr(1)
10170 .nr(16)
10171 .kr(4)
10172 .sr(1)
10173 .m(m)
10174 .n(n)
10175 .k(k)
10176 .ks(3)
10177 .iterations(1)
10178 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10179 }
10180 }
10181 }
10182 }
10183
10184 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_gt_16_small_kernel) {
10185 TEST_REQUIRES_ARM_NEON;
10186 for (uint32_t n = 17; n < 32; n++) {
10187 for (size_t k = 1; k <= 80; k += 17) {
10188 GemmMicrokernelTester()
10189 .mr(1)
10190 .nr(16)
10191 .kr(4)
10192 .sr(1)
10193 .m(1)
10194 .n(n)
10195 .k(k)
10196 .ks(3)
10197 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10198 }
10199 }
10200 }
10201
10202 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, n_div_16_small_kernel) {
10203 TEST_REQUIRES_ARM_NEON;
10204 for (uint32_t n = 32; n <= 48; n += 16) {
10205 for (size_t k = 1; k <= 80; k += 17) {
10206 GemmMicrokernelTester()
10207 .mr(1)
10208 .nr(16)
10209 .kr(4)
10210 .sr(1)
10211 .m(1)
10212 .n(n)
10213 .k(k)
10214 .ks(3)
10215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10216 }
10217 }
10218 }
10219
10220 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, strided_cm_subtile) {
10221 TEST_REQUIRES_ARM_NEON;
10222 for (size_t k = 1; k <= 80; k += 17) {
10223 for (uint32_t n = 1; n <= 16; n++) {
10224 for (uint32_t m = 1; m <= 1; m++) {
10225 GemmMicrokernelTester()
10226 .mr(1)
10227 .nr(16)
10228 .kr(4)
10229 .sr(1)
10230 .m(m)
10231 .n(n)
10232 .k(k)
10233 .cm_stride(19)
10234 .iterations(1)
10235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10236 }
10237 }
10238 }
10239 }
10240
10241 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, a_offset) {
10242 TEST_REQUIRES_ARM_NEON;
10243 for (size_t k = 1; k <= 80; k += 17) {
10244 GemmMicrokernelTester()
10245 .mr(1)
10246 .nr(16)
10247 .kr(4)
10248 .sr(1)
10249 .m(1)
10250 .n(16)
10251 .k(k)
10252 .ks(3)
10253 .a_offset(83)
10254 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10255 }
10256 }
10257
10258 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, zero) {
10259 TEST_REQUIRES_ARM_NEON;
10260 for (size_t k = 1; k <= 80; k += 17) {
10261 for (uint32_t mz = 0; mz < 1; mz++) {
10262 GemmMicrokernelTester()
10263 .mr(1)
10264 .nr(16)
10265 .kr(4)
10266 .sr(1)
10267 .m(1)
10268 .n(16)
10269 .k(k)
10270 .ks(3)
10271 .a_offset(83)
10272 .zero_index(mz)
10273 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10274 }
10275 }
10276 }
10277
10278 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, qmin) {
10279 TEST_REQUIRES_ARM_NEON;
10280 GemmMicrokernelTester()
10281 .mr(1)
10282 .nr(16)
10283 .kr(4)
10284 .sr(1)
10285 .m(1)
10286 .n(16)
10287 .k(16)
10288 .qmin(128)
10289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10290 }
10291
10292 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, qmax) {
10293 TEST_REQUIRES_ARM_NEON;
10294 GemmMicrokernelTester()
10295 .mr(1)
10296 .nr(16)
10297 .kr(4)
10298 .sr(1)
10299 .m(1)
10300 .n(16)
10301 .k(16)
10302 .qmax(128)
10303 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10304 }
10305
10306 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C4__NEON_MLAL_DUP, strided_cm) {
10307 TEST_REQUIRES_ARM_NEON;
10308 GemmMicrokernelTester()
10309 .mr(1)
10310 .nr(16)
10311 .kr(4)
10312 .sr(1)
10313 .m(1)
10314 .n(16)
10315 .k(16)
10316 .cm_stride(19)
10317 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c4__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10318 }
10319#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
10320
10321
10322#if XNN_ARCH_ARM || XNN_ARCH_ARM64
10323 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8) {
10324 TEST_REQUIRES_ARM_NEON;
10325 GemmMicrokernelTester()
10326 .mr(1)
10327 .nr(8)
10328 .kr(4)
10329 .sr(1)
10330 .m(1)
10331 .n(8)
10332 .k(8)
10333 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10334 }
10335
10336 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, strided_cn) {
10337 TEST_REQUIRES_ARM_NEON;
10338 GemmMicrokernelTester()
10339 .mr(1)
10340 .nr(8)
10341 .kr(4)
10342 .sr(1)
10343 .m(1)
10344 .n(8)
10345 .k(8)
10346 .cn_stride(11)
10347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10348 }
10349
10350 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8_subtile) {
10351 TEST_REQUIRES_ARM_NEON;
10352 for (uint32_t n = 1; n <= 8; n++) {
10353 for (uint32_t m = 1; m <= 1; m++) {
10354 GemmMicrokernelTester()
10355 .mr(1)
10356 .nr(8)
10357 .kr(4)
10358 .sr(1)
10359 .m(m)
10360 .n(n)
10361 .k(8)
10362 .iterations(1)
10363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10364 }
10365 }
10366 }
10367
10368 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8_subtile_m) {
10369 TEST_REQUIRES_ARM_NEON;
10370 for (uint32_t m = 1; m <= 1; m++) {
10371 GemmMicrokernelTester()
10372 .mr(1)
10373 .nr(8)
10374 .kr(4)
10375 .sr(1)
10376 .m(m)
10377 .n(8)
10378 .k(8)
10379 .iterations(1)
10380 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10381 }
10382 }
10383
10384 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_eq_8_subtile_n) {
10385 TEST_REQUIRES_ARM_NEON;
10386 for (uint32_t n = 1; n <= 8; n++) {
10387 GemmMicrokernelTester()
10388 .mr(1)
10389 .nr(8)
10390 .kr(4)
10391 .sr(1)
10392 .m(1)
10393 .n(n)
10394 .k(8)
10395 .iterations(1)
10396 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10397 }
10398 }
10399
10400 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_lt_8) {
10401 TEST_REQUIRES_ARM_NEON;
10402 for (size_t k = 1; k < 8; k++) {
10403 GemmMicrokernelTester()
10404 .mr(1)
10405 .nr(8)
10406 .kr(4)
10407 .sr(1)
10408 .m(1)
10409 .n(8)
10410 .k(k)
10411 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10412 }
10413 }
10414
10415 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_lt_8_subtile) {
10416 TEST_REQUIRES_ARM_NEON;
10417 for (size_t k = 1; k < 8; k++) {
10418 for (uint32_t n = 1; n <= 8; n++) {
10419 for (uint32_t m = 1; m <= 1; m++) {
10420 GemmMicrokernelTester()
10421 .mr(1)
10422 .nr(8)
10423 .kr(4)
10424 .sr(1)
10425 .m(m)
10426 .n(n)
10427 .k(k)
10428 .iterations(1)
10429 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10430 }
10431 }
10432 }
10433 }
10434
10435 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_gt_8) {
10436 TEST_REQUIRES_ARM_NEON;
10437 for (size_t k = 9; k < 16; k++) {
10438 GemmMicrokernelTester()
10439 .mr(1)
10440 .nr(8)
10441 .kr(4)
10442 .sr(1)
10443 .m(1)
10444 .n(8)
10445 .k(k)
10446 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10447 }
10448 }
10449
10450 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_gt_8_subtile) {
10451 TEST_REQUIRES_ARM_NEON;
10452 for (size_t k = 9; k < 16; k++) {
10453 for (uint32_t n = 1; n <= 8; n++) {
10454 for (uint32_t m = 1; m <= 1; m++) {
10455 GemmMicrokernelTester()
10456 .mr(1)
10457 .nr(8)
10458 .kr(4)
10459 .sr(1)
10460 .m(m)
10461 .n(n)
10462 .k(k)
10463 .iterations(1)
10464 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10465 }
10466 }
10467 }
10468 }
10469
10470 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_div_8) {
10471 TEST_REQUIRES_ARM_NEON;
10472 for (size_t k = 16; k <= 80; k += 8) {
10473 GemmMicrokernelTester()
10474 .mr(1)
10475 .nr(8)
10476 .kr(4)
10477 .sr(1)
10478 .m(1)
10479 .n(8)
10480 .k(k)
10481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10482 }
10483 }
10484
10485 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, k_div_8_subtile) {
10486 TEST_REQUIRES_ARM_NEON;
10487 for (size_t k = 16; k <= 80; k += 8) {
10488 for (uint32_t n = 1; n <= 8; n++) {
10489 for (uint32_t m = 1; m <= 1; m++) {
10490 GemmMicrokernelTester()
10491 .mr(1)
10492 .nr(8)
10493 .kr(4)
10494 .sr(1)
10495 .m(m)
10496 .n(n)
10497 .k(k)
10498 .iterations(1)
10499 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10500 }
10501 }
10502 }
10503 }
10504
10505 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8) {
10506 TEST_REQUIRES_ARM_NEON;
10507 for (uint32_t n = 9; n < 16; n++) {
10508 for (size_t k = 1; k <= 40; k += 9) {
10509 GemmMicrokernelTester()
10510 .mr(1)
10511 .nr(8)
10512 .kr(4)
10513 .sr(1)
10514 .m(1)
10515 .n(n)
10516 .k(k)
10517 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10518 }
10519 }
10520 }
10521
10522 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8_strided_cn) {
10523 TEST_REQUIRES_ARM_NEON;
10524 for (uint32_t n = 9; n < 16; n++) {
10525 for (size_t k = 1; k <= 40; k += 9) {
10526 GemmMicrokernelTester()
10527 .mr(1)
10528 .nr(8)
10529 .kr(4)
10530 .sr(1)
10531 .m(1)
10532 .n(n)
10533 .k(k)
10534 .cn_stride(11)
10535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10536 }
10537 }
10538 }
10539
10540 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8_subtile) {
10541 TEST_REQUIRES_ARM_NEON;
10542 for (uint32_t n = 9; n < 16; n++) {
10543 for (size_t k = 1; k <= 40; k += 9) {
10544 for (uint32_t m = 1; m <= 1; m++) {
10545 GemmMicrokernelTester()
10546 .mr(1)
10547 .nr(8)
10548 .kr(4)
10549 .sr(1)
10550 .m(m)
10551 .n(n)
10552 .k(k)
10553 .iterations(1)
10554 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10555 }
10556 }
10557 }
10558 }
10559
10560 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8) {
10561 TEST_REQUIRES_ARM_NEON;
10562 for (uint32_t n = 16; n <= 24; n += 8) {
10563 for (size_t k = 1; k <= 40; k += 9) {
10564 GemmMicrokernelTester()
10565 .mr(1)
10566 .nr(8)
10567 .kr(4)
10568 .sr(1)
10569 .m(1)
10570 .n(n)
10571 .k(k)
10572 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10573 }
10574 }
10575 }
10576
10577 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8_strided_cn) {
10578 TEST_REQUIRES_ARM_NEON;
10579 for (uint32_t n = 16; n <= 24; n += 8) {
10580 for (size_t k = 1; k <= 40; k += 9) {
10581 GemmMicrokernelTester()
10582 .mr(1)
10583 .nr(8)
10584 .kr(4)
10585 .sr(1)
10586 .m(1)
10587 .n(n)
10588 .k(k)
10589 .cn_stride(11)
10590 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10591 }
10592 }
10593 }
10594
10595 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8_subtile) {
10596 TEST_REQUIRES_ARM_NEON;
10597 for (uint32_t n = 16; n <= 24; n += 8) {
10598 for (size_t k = 1; k <= 40; k += 9) {
10599 for (uint32_t m = 1; m <= 1; m++) {
10600 GemmMicrokernelTester()
10601 .mr(1)
10602 .nr(8)
10603 .kr(4)
10604 .sr(1)
10605 .m(m)
10606 .n(n)
10607 .k(k)
10608 .iterations(1)
10609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10610 }
10611 }
10612 }
10613 }
10614
10615 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, small_kernel) {
10616 TEST_REQUIRES_ARM_NEON;
10617 for (size_t k = 1; k <= 40; k += 9) {
10618 GemmMicrokernelTester()
10619 .mr(1)
10620 .nr(8)
10621 .kr(4)
10622 .sr(1)
10623 .m(1)
10624 .n(8)
10625 .k(k)
10626 .ks(3)
10627 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10628 }
10629 }
10630
10631 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, small_kernel_subtile) {
10632 TEST_REQUIRES_ARM_NEON;
10633 for (size_t k = 1; k <= 40; k += 9) {
10634 for (uint32_t n = 1; n <= 8; n++) {
10635 for (uint32_t m = 1; m <= 1; m++) {
10636 GemmMicrokernelTester()
10637 .mr(1)
10638 .nr(8)
10639 .kr(4)
10640 .sr(1)
10641 .m(m)
10642 .n(n)
10643 .k(k)
10644 .ks(3)
10645 .iterations(1)
10646 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10647 }
10648 }
10649 }
10650 }
10651
10652 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_gt_8_small_kernel) {
10653 TEST_REQUIRES_ARM_NEON;
10654 for (uint32_t n = 9; n < 16; n++) {
10655 for (size_t k = 1; k <= 40; k += 9) {
10656 GemmMicrokernelTester()
10657 .mr(1)
10658 .nr(8)
10659 .kr(4)
10660 .sr(1)
10661 .m(1)
10662 .n(n)
10663 .k(k)
10664 .ks(3)
10665 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10666 }
10667 }
10668 }
10669
10670 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, n_div_8_small_kernel) {
10671 TEST_REQUIRES_ARM_NEON;
10672 for (uint32_t n = 16; n <= 24; n += 8) {
10673 for (size_t k = 1; k <= 40; k += 9) {
10674 GemmMicrokernelTester()
10675 .mr(1)
10676 .nr(8)
10677 .kr(4)
10678 .sr(1)
10679 .m(1)
10680 .n(n)
10681 .k(k)
10682 .ks(3)
10683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10684 }
10685 }
10686 }
10687
10688 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, strided_cm_subtile) {
10689 TEST_REQUIRES_ARM_NEON;
10690 for (size_t k = 1; k <= 40; k += 9) {
10691 for (uint32_t n = 1; n <= 8; n++) {
10692 for (uint32_t m = 1; m <= 1; m++) {
10693 GemmMicrokernelTester()
10694 .mr(1)
10695 .nr(8)
10696 .kr(4)
10697 .sr(1)
10698 .m(m)
10699 .n(n)
10700 .k(k)
10701 .cm_stride(11)
10702 .iterations(1)
10703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10704 }
10705 }
10706 }
10707 }
10708
10709 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, a_offset) {
10710 TEST_REQUIRES_ARM_NEON;
10711 for (size_t k = 1; k <= 40; k += 9) {
10712 GemmMicrokernelTester()
10713 .mr(1)
10714 .nr(8)
10715 .kr(4)
10716 .sr(1)
10717 .m(1)
10718 .n(8)
10719 .k(k)
10720 .ks(3)
10721 .a_offset(43)
10722 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10723 }
10724 }
10725
10726 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, zero) {
10727 TEST_REQUIRES_ARM_NEON;
10728 for (size_t k = 1; k <= 40; k += 9) {
10729 for (uint32_t mz = 0; mz < 1; mz++) {
10730 GemmMicrokernelTester()
10731 .mr(1)
10732 .nr(8)
10733 .kr(4)
10734 .sr(1)
10735 .m(1)
10736 .n(8)
10737 .k(k)
10738 .ks(3)
10739 .a_offset(43)
10740 .zero_index(mz)
10741 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10742 }
10743 }
10744 }
10745
10746 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, qmin) {
10747 TEST_REQUIRES_ARM_NEON;
10748 GemmMicrokernelTester()
10749 .mr(1)
10750 .nr(8)
10751 .kr(4)
10752 .sr(1)
10753 .m(1)
10754 .n(8)
10755 .k(8)
10756 .qmin(128)
10757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10758 }
10759
10760 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, qmax) {
10761 TEST_REQUIRES_ARM_NEON;
10762 GemmMicrokernelTester()
10763 .mr(1)
10764 .nr(8)
10765 .kr(4)
10766 .sr(1)
10767 .m(1)
10768 .n(8)
10769 .k(8)
10770 .qmax(128)
10771 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10772 }
10773
10774 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MULL_LD1R, strided_cm) {
10775 TEST_REQUIRES_ARM_NEON;
10776 GemmMicrokernelTester()
10777 .mr(1)
10778 .nr(8)
10779 .kr(4)
10780 .sr(1)
10781 .m(1)
10782 .n(8)
10783 .k(8)
10784 .cm_stride(11)
10785 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mull_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10786 }
10787#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
10788
10789
10790#if XNN_ARCH_ARM || XNN_ARCH_ARM64
10791 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16) {
10792 TEST_REQUIRES_ARM_NEON;
10793 GemmMicrokernelTester()
10794 .mr(1)
10795 .nr(8)
10796 .kr(4)
10797 .sr(1)
10798 .m(1)
10799 .n(8)
10800 .k(16)
10801 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10802 }
10803
10804 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, strided_cn) {
10805 TEST_REQUIRES_ARM_NEON;
10806 GemmMicrokernelTester()
10807 .mr(1)
10808 .nr(8)
10809 .kr(4)
10810 .sr(1)
10811 .m(1)
10812 .n(8)
10813 .k(16)
10814 .cn_stride(11)
10815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10816 }
10817
10818 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16_subtile) {
10819 TEST_REQUIRES_ARM_NEON;
10820 for (uint32_t n = 1; n <= 8; n++) {
10821 for (uint32_t m = 1; m <= 1; m++) {
10822 GemmMicrokernelTester()
10823 .mr(1)
10824 .nr(8)
10825 .kr(4)
10826 .sr(1)
10827 .m(m)
10828 .n(n)
10829 .k(16)
10830 .iterations(1)
10831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10832 }
10833 }
10834 }
10835
10836 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16_subtile_m) {
10837 TEST_REQUIRES_ARM_NEON;
10838 for (uint32_t m = 1; m <= 1; m++) {
10839 GemmMicrokernelTester()
10840 .mr(1)
10841 .nr(8)
10842 .kr(4)
10843 .sr(1)
10844 .m(m)
10845 .n(8)
10846 .k(16)
10847 .iterations(1)
10848 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10849 }
10850 }
10851
10852 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_eq_16_subtile_n) {
10853 TEST_REQUIRES_ARM_NEON;
10854 for (uint32_t n = 1; n <= 8; n++) {
10855 GemmMicrokernelTester()
10856 .mr(1)
10857 .nr(8)
10858 .kr(4)
10859 .sr(1)
10860 .m(1)
10861 .n(n)
10862 .k(16)
10863 .iterations(1)
10864 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10865 }
10866 }
10867
10868 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_lt_16) {
10869 TEST_REQUIRES_ARM_NEON;
10870 for (size_t k = 1; k < 16; k++) {
10871 GemmMicrokernelTester()
10872 .mr(1)
10873 .nr(8)
10874 .kr(4)
10875 .sr(1)
10876 .m(1)
10877 .n(8)
10878 .k(k)
10879 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10880 }
10881 }
10882
10883 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_lt_16_subtile) {
10884 TEST_REQUIRES_ARM_NEON;
10885 for (size_t k = 1; k < 16; k++) {
10886 for (uint32_t n = 1; n <= 8; n++) {
10887 for (uint32_t m = 1; m <= 1; m++) {
10888 GemmMicrokernelTester()
10889 .mr(1)
10890 .nr(8)
10891 .kr(4)
10892 .sr(1)
10893 .m(m)
10894 .n(n)
10895 .k(k)
10896 .iterations(1)
10897 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10898 }
10899 }
10900 }
10901 }
10902
10903 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_gt_16) {
10904 TEST_REQUIRES_ARM_NEON;
10905 for (size_t k = 17; k < 32; k++) {
10906 GemmMicrokernelTester()
10907 .mr(1)
10908 .nr(8)
10909 .kr(4)
10910 .sr(1)
10911 .m(1)
10912 .n(8)
10913 .k(k)
10914 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10915 }
10916 }
10917
10918 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_gt_16_subtile) {
10919 TEST_REQUIRES_ARM_NEON;
10920 for (size_t k = 17; k < 32; k++) {
10921 for (uint32_t n = 1; n <= 8; n++) {
10922 for (uint32_t m = 1; m <= 1; m++) {
10923 GemmMicrokernelTester()
10924 .mr(1)
10925 .nr(8)
10926 .kr(4)
10927 .sr(1)
10928 .m(m)
10929 .n(n)
10930 .k(k)
10931 .iterations(1)
10932 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10933 }
10934 }
10935 }
10936 }
10937
10938 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_div_16) {
10939 TEST_REQUIRES_ARM_NEON;
10940 for (size_t k = 32; k <= 160; k += 16) {
10941 GemmMicrokernelTester()
10942 .mr(1)
10943 .nr(8)
10944 .kr(4)
10945 .sr(1)
10946 .m(1)
10947 .n(8)
10948 .k(k)
10949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10950 }
10951 }
10952
10953 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, k_div_16_subtile) {
10954 TEST_REQUIRES_ARM_NEON;
10955 for (size_t k = 32; k <= 160; k += 16) {
10956 for (uint32_t n = 1; n <= 8; n++) {
10957 for (uint32_t m = 1; m <= 1; m++) {
10958 GemmMicrokernelTester()
10959 .mr(1)
10960 .nr(8)
10961 .kr(4)
10962 .sr(1)
10963 .m(m)
10964 .n(n)
10965 .k(k)
10966 .iterations(1)
10967 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10968 }
10969 }
10970 }
10971 }
10972
10973 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8) {
10974 TEST_REQUIRES_ARM_NEON;
10975 for (uint32_t n = 9; n < 16; n++) {
10976 for (size_t k = 1; k <= 80; k += 17) {
10977 GemmMicrokernelTester()
10978 .mr(1)
10979 .nr(8)
10980 .kr(4)
10981 .sr(1)
10982 .m(1)
10983 .n(n)
10984 .k(k)
10985 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
10986 }
10987 }
10988 }
10989
10990 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8_strided_cn) {
10991 TEST_REQUIRES_ARM_NEON;
10992 for (uint32_t n = 9; n < 16; n++) {
10993 for (size_t k = 1; k <= 80; k += 17) {
10994 GemmMicrokernelTester()
10995 .mr(1)
10996 .nr(8)
10997 .kr(4)
10998 .sr(1)
10999 .m(1)
11000 .n(n)
11001 .k(k)
11002 .cn_stride(11)
11003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11004 }
11005 }
11006 }
11007
11008 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8_subtile) {
11009 TEST_REQUIRES_ARM_NEON;
11010 for (uint32_t n = 9; n < 16; n++) {
11011 for (size_t k = 1; k <= 80; k += 17) {
11012 for (uint32_t m = 1; m <= 1; m++) {
11013 GemmMicrokernelTester()
11014 .mr(1)
11015 .nr(8)
11016 .kr(4)
11017 .sr(1)
11018 .m(m)
11019 .n(n)
11020 .k(k)
11021 .iterations(1)
11022 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11023 }
11024 }
11025 }
11026 }
11027
11028 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8) {
11029 TEST_REQUIRES_ARM_NEON;
11030 for (uint32_t n = 16; n <= 24; n += 8) {
11031 for (size_t k = 1; k <= 80; k += 17) {
11032 GemmMicrokernelTester()
11033 .mr(1)
11034 .nr(8)
11035 .kr(4)
11036 .sr(1)
11037 .m(1)
11038 .n(n)
11039 .k(k)
11040 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11041 }
11042 }
11043 }
11044
11045 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8_strided_cn) {
11046 TEST_REQUIRES_ARM_NEON;
11047 for (uint32_t n = 16; n <= 24; n += 8) {
11048 for (size_t k = 1; k <= 80; k += 17) {
11049 GemmMicrokernelTester()
11050 .mr(1)
11051 .nr(8)
11052 .kr(4)
11053 .sr(1)
11054 .m(1)
11055 .n(n)
11056 .k(k)
11057 .cn_stride(11)
11058 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11059 }
11060 }
11061 }
11062
11063 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8_subtile) {
11064 TEST_REQUIRES_ARM_NEON;
11065 for (uint32_t n = 16; n <= 24; n += 8) {
11066 for (size_t k = 1; k <= 80; k += 17) {
11067 for (uint32_t m = 1; m <= 1; m++) {
11068 GemmMicrokernelTester()
11069 .mr(1)
11070 .nr(8)
11071 .kr(4)
11072 .sr(1)
11073 .m(m)
11074 .n(n)
11075 .k(k)
11076 .iterations(1)
11077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11078 }
11079 }
11080 }
11081 }
11082
11083 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, small_kernel) {
11084 TEST_REQUIRES_ARM_NEON;
11085 for (size_t k = 1; k <= 80; k += 17) {
11086 GemmMicrokernelTester()
11087 .mr(1)
11088 .nr(8)
11089 .kr(4)
11090 .sr(1)
11091 .m(1)
11092 .n(8)
11093 .k(k)
11094 .ks(3)
11095 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11096 }
11097 }
11098
11099 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, small_kernel_subtile) {
11100 TEST_REQUIRES_ARM_NEON;
11101 for (size_t k = 1; k <= 80; k += 17) {
11102 for (uint32_t n = 1; n <= 8; n++) {
11103 for (uint32_t m = 1; m <= 1; m++) {
11104 GemmMicrokernelTester()
11105 .mr(1)
11106 .nr(8)
11107 .kr(4)
11108 .sr(1)
11109 .m(m)
11110 .n(n)
11111 .k(k)
11112 .ks(3)
11113 .iterations(1)
11114 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11115 }
11116 }
11117 }
11118 }
11119
11120 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_gt_8_small_kernel) {
11121 TEST_REQUIRES_ARM_NEON;
11122 for (uint32_t n = 9; n < 16; n++) {
11123 for (size_t k = 1; k <= 80; k += 17) {
11124 GemmMicrokernelTester()
11125 .mr(1)
11126 .nr(8)
11127 .kr(4)
11128 .sr(1)
11129 .m(1)
11130 .n(n)
11131 .k(k)
11132 .ks(3)
11133 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11134 }
11135 }
11136 }
11137
11138 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, n_div_8_small_kernel) {
11139 TEST_REQUIRES_ARM_NEON;
11140 for (uint32_t n = 16; n <= 24; n += 8) {
11141 for (size_t k = 1; k <= 80; k += 17) {
11142 GemmMicrokernelTester()
11143 .mr(1)
11144 .nr(8)
11145 .kr(4)
11146 .sr(1)
11147 .m(1)
11148 .n(n)
11149 .k(k)
11150 .ks(3)
11151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11152 }
11153 }
11154 }
11155
11156 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, strided_cm_subtile) {
11157 TEST_REQUIRES_ARM_NEON;
11158 for (size_t k = 1; k <= 80; k += 17) {
11159 for (uint32_t n = 1; n <= 8; n++) {
11160 for (uint32_t m = 1; m <= 1; m++) {
11161 GemmMicrokernelTester()
11162 .mr(1)
11163 .nr(8)
11164 .kr(4)
11165 .sr(1)
11166 .m(m)
11167 .n(n)
11168 .k(k)
11169 .cm_stride(11)
11170 .iterations(1)
11171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11172 }
11173 }
11174 }
11175 }
11176
11177 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, a_offset) {
11178 TEST_REQUIRES_ARM_NEON;
11179 for (size_t k = 1; k <= 80; k += 17) {
11180 GemmMicrokernelTester()
11181 .mr(1)
11182 .nr(8)
11183 .kr(4)
11184 .sr(1)
11185 .m(1)
11186 .n(8)
11187 .k(k)
11188 .ks(3)
11189 .a_offset(83)
11190 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11191 }
11192 }
11193
11194 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, zero) {
11195 TEST_REQUIRES_ARM_NEON;
11196 for (size_t k = 1; k <= 80; k += 17) {
11197 for (uint32_t mz = 0; mz < 1; mz++) {
11198 GemmMicrokernelTester()
11199 .mr(1)
11200 .nr(8)
11201 .kr(4)
11202 .sr(1)
11203 .m(1)
11204 .n(8)
11205 .k(k)
11206 .ks(3)
11207 .a_offset(83)
11208 .zero_index(mz)
11209 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11210 }
11211 }
11212 }
11213
11214 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, qmin) {
11215 TEST_REQUIRES_ARM_NEON;
11216 GemmMicrokernelTester()
11217 .mr(1)
11218 .nr(8)
11219 .kr(4)
11220 .sr(1)
11221 .m(1)
11222 .n(8)
11223 .k(16)
11224 .qmin(128)
11225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11226 }
11227
11228 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, qmax) {
11229 TEST_REQUIRES_ARM_NEON;
11230 GemmMicrokernelTester()
11231 .mr(1)
11232 .nr(8)
11233 .kr(4)
11234 .sr(1)
11235 .m(1)
11236 .n(8)
11237 .k(16)
11238 .qmax(128)
11239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11240 }
11241
11242 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C4__NEON_MLAL_LD1R, strided_cm) {
11243 TEST_REQUIRES_ARM_NEON;
11244 GemmMicrokernelTester()
11245 .mr(1)
11246 .nr(8)
11247 .kr(4)
11248 .sr(1)
11249 .m(1)
11250 .n(8)
11251 .k(16)
11252 .cm_stride(11)
11253 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c4__neon_mlal_ld1r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11254 }
11255#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
11256
11257
11258#if XNN_ARCH_ARM || XNN_ARCH_ARM64
11259 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_eq_8) {
11260 TEST_REQUIRES_ARM_NEON;
11261 GemmMicrokernelTester()
11262 .mr(2)
11263 .nr(8)
11264 .kr(4)
11265 .sr(1)
11266 .m(2)
11267 .n(8)
11268 .k(8)
11269 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11270 }
11271
11272 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, strided_cn) {
11273 TEST_REQUIRES_ARM_NEON;
11274 GemmMicrokernelTester()
11275 .mr(2)
11276 .nr(8)
11277 .kr(4)
11278 .sr(1)
11279 .m(2)
11280 .n(8)
11281 .k(8)
11282 .cn_stride(11)
11283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11284 }
11285
11286 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_eq_8_subtile) {
11287 TEST_REQUIRES_ARM_NEON;
11288 for (uint32_t n = 1; n <= 8; n++) {
11289 for (uint32_t m = 1; m <= 2; m++) {
11290 GemmMicrokernelTester()
11291 .mr(2)
11292 .nr(8)
11293 .kr(4)
11294 .sr(1)
11295 .m(m)
11296 .n(n)
11297 .k(8)
11298 .iterations(1)
11299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11300 }
11301 }
11302 }
11303
11304 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_eq_8_subtile_m) {
11305 TEST_REQUIRES_ARM_NEON;
11306 for (uint32_t m = 1; m <= 2; m++) {
11307 GemmMicrokernelTester()
11308 .mr(2)
11309 .nr(8)
11310 .kr(4)
11311 .sr(1)
11312 .m(m)
11313 .n(8)
11314 .k(8)
11315 .iterations(1)
11316 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11317 }
11318 }
11319
11320 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_eq_8_subtile_n) {
11321 TEST_REQUIRES_ARM_NEON;
11322 for (uint32_t n = 1; n <= 8; n++) {
11323 GemmMicrokernelTester()
11324 .mr(2)
11325 .nr(8)
11326 .kr(4)
11327 .sr(1)
11328 .m(2)
11329 .n(n)
11330 .k(8)
11331 .iterations(1)
11332 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11333 }
11334 }
11335
11336 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_lt_8) {
11337 TEST_REQUIRES_ARM_NEON;
11338 for (size_t k = 1; k < 8; k++) {
11339 GemmMicrokernelTester()
11340 .mr(2)
11341 .nr(8)
11342 .kr(4)
11343 .sr(1)
11344 .m(2)
11345 .n(8)
11346 .k(k)
11347 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11348 }
11349 }
11350
11351 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_lt_8_subtile) {
11352 TEST_REQUIRES_ARM_NEON;
11353 for (size_t k = 1; k < 8; k++) {
11354 for (uint32_t n = 1; n <= 8; n++) {
11355 for (uint32_t m = 1; m <= 2; m++) {
11356 GemmMicrokernelTester()
11357 .mr(2)
11358 .nr(8)
11359 .kr(4)
11360 .sr(1)
11361 .m(m)
11362 .n(n)
11363 .k(k)
11364 .iterations(1)
11365 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11366 }
11367 }
11368 }
11369 }
11370
11371 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_gt_8) {
11372 TEST_REQUIRES_ARM_NEON;
11373 for (size_t k = 9; k < 16; k++) {
11374 GemmMicrokernelTester()
11375 .mr(2)
11376 .nr(8)
11377 .kr(4)
11378 .sr(1)
11379 .m(2)
11380 .n(8)
11381 .k(k)
11382 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11383 }
11384 }
11385
11386 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_gt_8_subtile) {
11387 TEST_REQUIRES_ARM_NEON;
11388 for (size_t k = 9; k < 16; k++) {
11389 for (uint32_t n = 1; n <= 8; n++) {
11390 for (uint32_t m = 1; m <= 2; m++) {
11391 GemmMicrokernelTester()
11392 .mr(2)
11393 .nr(8)
11394 .kr(4)
11395 .sr(1)
11396 .m(m)
11397 .n(n)
11398 .k(k)
11399 .iterations(1)
11400 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11401 }
11402 }
11403 }
11404 }
11405
11406 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_div_8) {
11407 TEST_REQUIRES_ARM_NEON;
11408 for (size_t k = 16; k <= 80; k += 8) {
11409 GemmMicrokernelTester()
11410 .mr(2)
11411 .nr(8)
11412 .kr(4)
11413 .sr(1)
11414 .m(2)
11415 .n(8)
11416 .k(k)
11417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11418 }
11419 }
11420
11421 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, k_div_8_subtile) {
11422 TEST_REQUIRES_ARM_NEON;
11423 for (size_t k = 16; k <= 80; k += 8) {
11424 for (uint32_t n = 1; n <= 8; n++) {
11425 for (uint32_t m = 1; m <= 2; m++) {
11426 GemmMicrokernelTester()
11427 .mr(2)
11428 .nr(8)
11429 .kr(4)
11430 .sr(1)
11431 .m(m)
11432 .n(n)
11433 .k(k)
11434 .iterations(1)
11435 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11436 }
11437 }
11438 }
11439 }
11440
11441 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_gt_8) {
11442 TEST_REQUIRES_ARM_NEON;
11443 for (uint32_t n = 9; n < 16; n++) {
11444 for (size_t k = 1; k <= 40; k += 9) {
11445 GemmMicrokernelTester()
11446 .mr(2)
11447 .nr(8)
11448 .kr(4)
11449 .sr(1)
11450 .m(2)
11451 .n(n)
11452 .k(k)
11453 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11454 }
11455 }
11456 }
11457
11458 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_gt_8_strided_cn) {
11459 TEST_REQUIRES_ARM_NEON;
11460 for (uint32_t n = 9; n < 16; n++) {
11461 for (size_t k = 1; k <= 40; k += 9) {
11462 GemmMicrokernelTester()
11463 .mr(2)
11464 .nr(8)
11465 .kr(4)
11466 .sr(1)
11467 .m(2)
11468 .n(n)
11469 .k(k)
11470 .cn_stride(11)
11471 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11472 }
11473 }
11474 }
11475
11476 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_gt_8_subtile) {
11477 TEST_REQUIRES_ARM_NEON;
11478 for (uint32_t n = 9; n < 16; n++) {
11479 for (size_t k = 1; k <= 40; k += 9) {
11480 for (uint32_t m = 1; m <= 2; m++) {
11481 GemmMicrokernelTester()
11482 .mr(2)
11483 .nr(8)
11484 .kr(4)
11485 .sr(1)
11486 .m(m)
11487 .n(n)
11488 .k(k)
11489 .iterations(1)
11490 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11491 }
11492 }
11493 }
11494 }
11495
11496 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_div_8) {
11497 TEST_REQUIRES_ARM_NEON;
11498 for (uint32_t n = 16; n <= 24; n += 8) {
11499 for (size_t k = 1; k <= 40; k += 9) {
11500 GemmMicrokernelTester()
11501 .mr(2)
11502 .nr(8)
11503 .kr(4)
11504 .sr(1)
11505 .m(2)
11506 .n(n)
11507 .k(k)
11508 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11509 }
11510 }
11511 }
11512
11513 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_div_8_strided_cn) {
11514 TEST_REQUIRES_ARM_NEON;
11515 for (uint32_t n = 16; n <= 24; n += 8) {
11516 for (size_t k = 1; k <= 40; k += 9) {
11517 GemmMicrokernelTester()
11518 .mr(2)
11519 .nr(8)
11520 .kr(4)
11521 .sr(1)
11522 .m(2)
11523 .n(n)
11524 .k(k)
11525 .cn_stride(11)
11526 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11527 }
11528 }
11529 }
11530
11531 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_div_8_subtile) {
11532 TEST_REQUIRES_ARM_NEON;
11533 for (uint32_t n = 16; n <= 24; n += 8) {
11534 for (size_t k = 1; k <= 40; k += 9) {
11535 for (uint32_t m = 1; m <= 2; m++) {
11536 GemmMicrokernelTester()
11537 .mr(2)
11538 .nr(8)
11539 .kr(4)
11540 .sr(1)
11541 .m(m)
11542 .n(n)
11543 .k(k)
11544 .iterations(1)
11545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11546 }
11547 }
11548 }
11549 }
11550
11551 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, small_kernel) {
11552 TEST_REQUIRES_ARM_NEON;
11553 for (size_t k = 1; k <= 40; k += 9) {
11554 GemmMicrokernelTester()
11555 .mr(2)
11556 .nr(8)
11557 .kr(4)
11558 .sr(1)
11559 .m(2)
11560 .n(8)
11561 .k(k)
11562 .ks(3)
11563 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11564 }
11565 }
11566
11567 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, small_kernel_subtile) {
11568 TEST_REQUIRES_ARM_NEON;
11569 for (size_t k = 1; k <= 40; k += 9) {
11570 for (uint32_t n = 1; n <= 8; n++) {
11571 for (uint32_t m = 1; m <= 2; m++) {
11572 GemmMicrokernelTester()
11573 .mr(2)
11574 .nr(8)
11575 .kr(4)
11576 .sr(1)
11577 .m(m)
11578 .n(n)
11579 .k(k)
11580 .ks(3)
11581 .iterations(1)
11582 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11583 }
11584 }
11585 }
11586 }
11587
11588 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_gt_8_small_kernel) {
11589 TEST_REQUIRES_ARM_NEON;
11590 for (uint32_t n = 9; n < 16; n++) {
11591 for (size_t k = 1; k <= 40; k += 9) {
11592 GemmMicrokernelTester()
11593 .mr(2)
11594 .nr(8)
11595 .kr(4)
11596 .sr(1)
11597 .m(2)
11598 .n(n)
11599 .k(k)
11600 .ks(3)
11601 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11602 }
11603 }
11604 }
11605
11606 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, n_div_8_small_kernel) {
11607 TEST_REQUIRES_ARM_NEON;
11608 for (uint32_t n = 16; n <= 24; n += 8) {
11609 for (size_t k = 1; k <= 40; k += 9) {
11610 GemmMicrokernelTester()
11611 .mr(2)
11612 .nr(8)
11613 .kr(4)
11614 .sr(1)
11615 .m(2)
11616 .n(n)
11617 .k(k)
11618 .ks(3)
11619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11620 }
11621 }
11622 }
11623
11624 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, strided_cm_subtile) {
11625 TEST_REQUIRES_ARM_NEON;
11626 for (size_t k = 1; k <= 40; k += 9) {
11627 for (uint32_t n = 1; n <= 8; n++) {
11628 for (uint32_t m = 1; m <= 2; m++) {
11629 GemmMicrokernelTester()
11630 .mr(2)
11631 .nr(8)
11632 .kr(4)
11633 .sr(1)
11634 .m(m)
11635 .n(n)
11636 .k(k)
11637 .cm_stride(11)
11638 .iterations(1)
11639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11640 }
11641 }
11642 }
11643 }
11644
11645 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, a_offset) {
11646 TEST_REQUIRES_ARM_NEON;
11647 for (size_t k = 1; k <= 40; k += 9) {
11648 GemmMicrokernelTester()
11649 .mr(2)
11650 .nr(8)
11651 .kr(4)
11652 .sr(1)
11653 .m(2)
11654 .n(8)
11655 .k(k)
11656 .ks(3)
11657 .a_offset(83)
11658 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11659 }
11660 }
11661
11662 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, zero) {
11663 TEST_REQUIRES_ARM_NEON;
11664 for (size_t k = 1; k <= 40; k += 9) {
11665 for (uint32_t mz = 0; mz < 2; mz++) {
11666 GemmMicrokernelTester()
11667 .mr(2)
11668 .nr(8)
11669 .kr(4)
11670 .sr(1)
11671 .m(2)
11672 .n(8)
11673 .k(k)
11674 .ks(3)
11675 .a_offset(83)
11676 .zero_index(mz)
11677 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11678 }
11679 }
11680 }
11681
11682 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, qmin) {
11683 TEST_REQUIRES_ARM_NEON;
11684 GemmMicrokernelTester()
11685 .mr(2)
11686 .nr(8)
11687 .kr(4)
11688 .sr(1)
11689 .m(2)
11690 .n(8)
11691 .k(8)
11692 .qmin(128)
11693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11694 }
11695
11696 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, qmax) {
11697 TEST_REQUIRES_ARM_NEON;
11698 GemmMicrokernelTester()
11699 .mr(2)
11700 .nr(8)
11701 .kr(4)
11702 .sr(1)
11703 .m(2)
11704 .n(8)
11705 .k(8)
11706 .qmax(128)
11707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11708 }
11709
11710 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MULL_LD2R, strided_cm) {
11711 TEST_REQUIRES_ARM_NEON;
11712 GemmMicrokernelTester()
11713 .mr(2)
11714 .nr(8)
11715 .kr(4)
11716 .sr(1)
11717 .m(2)
11718 .n(8)
11719 .k(8)
11720 .cm_stride(11)
11721 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mull_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11722 }
11723#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
11724
11725
11726#if XNN_ARCH_ARM || XNN_ARCH_ARM64
11727 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_eq_16) {
11728 TEST_REQUIRES_ARM_NEON;
11729 GemmMicrokernelTester()
11730 .mr(2)
11731 .nr(8)
11732 .kr(4)
11733 .sr(1)
11734 .m(2)
11735 .n(8)
11736 .k(16)
11737 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11738 }
11739
11740 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, strided_cn) {
11741 TEST_REQUIRES_ARM_NEON;
11742 GemmMicrokernelTester()
11743 .mr(2)
11744 .nr(8)
11745 .kr(4)
11746 .sr(1)
11747 .m(2)
11748 .n(8)
11749 .k(16)
11750 .cn_stride(11)
11751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11752 }
11753
11754 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_eq_16_subtile) {
11755 TEST_REQUIRES_ARM_NEON;
11756 for (uint32_t n = 1; n <= 8; n++) {
11757 for (uint32_t m = 1; m <= 2; m++) {
11758 GemmMicrokernelTester()
11759 .mr(2)
11760 .nr(8)
11761 .kr(4)
11762 .sr(1)
11763 .m(m)
11764 .n(n)
11765 .k(16)
11766 .iterations(1)
11767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11768 }
11769 }
11770 }
11771
11772 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_eq_16_subtile_m) {
11773 TEST_REQUIRES_ARM_NEON;
11774 for (uint32_t m = 1; m <= 2; m++) {
11775 GemmMicrokernelTester()
11776 .mr(2)
11777 .nr(8)
11778 .kr(4)
11779 .sr(1)
11780 .m(m)
11781 .n(8)
11782 .k(16)
11783 .iterations(1)
11784 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11785 }
11786 }
11787
11788 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_eq_16_subtile_n) {
11789 TEST_REQUIRES_ARM_NEON;
11790 for (uint32_t n = 1; n <= 8; n++) {
11791 GemmMicrokernelTester()
11792 .mr(2)
11793 .nr(8)
11794 .kr(4)
11795 .sr(1)
11796 .m(2)
11797 .n(n)
11798 .k(16)
11799 .iterations(1)
11800 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11801 }
11802 }
11803
11804 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_lt_16) {
11805 TEST_REQUIRES_ARM_NEON;
11806 for (size_t k = 1; k < 16; k++) {
11807 GemmMicrokernelTester()
11808 .mr(2)
11809 .nr(8)
11810 .kr(4)
11811 .sr(1)
11812 .m(2)
11813 .n(8)
11814 .k(k)
11815 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11816 }
11817 }
11818
11819 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_lt_16_subtile) {
11820 TEST_REQUIRES_ARM_NEON;
11821 for (size_t k = 1; k < 16; k++) {
11822 for (uint32_t n = 1; n <= 8; n++) {
11823 for (uint32_t m = 1; m <= 2; m++) {
11824 GemmMicrokernelTester()
11825 .mr(2)
11826 .nr(8)
11827 .kr(4)
11828 .sr(1)
11829 .m(m)
11830 .n(n)
11831 .k(k)
11832 .iterations(1)
11833 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11834 }
11835 }
11836 }
11837 }
11838
11839 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_gt_16) {
11840 TEST_REQUIRES_ARM_NEON;
11841 for (size_t k = 17; k < 32; k++) {
11842 GemmMicrokernelTester()
11843 .mr(2)
11844 .nr(8)
11845 .kr(4)
11846 .sr(1)
11847 .m(2)
11848 .n(8)
11849 .k(k)
11850 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11851 }
11852 }
11853
11854 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_gt_16_subtile) {
11855 TEST_REQUIRES_ARM_NEON;
11856 for (size_t k = 17; k < 32; k++) {
11857 for (uint32_t n = 1; n <= 8; n++) {
11858 for (uint32_t m = 1; m <= 2; m++) {
11859 GemmMicrokernelTester()
11860 .mr(2)
11861 .nr(8)
11862 .kr(4)
11863 .sr(1)
11864 .m(m)
11865 .n(n)
11866 .k(k)
11867 .iterations(1)
11868 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11869 }
11870 }
11871 }
11872 }
11873
11874 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_div_16) {
11875 TEST_REQUIRES_ARM_NEON;
11876 for (size_t k = 32; k <= 160; k += 16) {
11877 GemmMicrokernelTester()
11878 .mr(2)
11879 .nr(8)
11880 .kr(4)
11881 .sr(1)
11882 .m(2)
11883 .n(8)
11884 .k(k)
11885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11886 }
11887 }
11888
11889 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, k_div_16_subtile) {
11890 TEST_REQUIRES_ARM_NEON;
11891 for (size_t k = 32; k <= 160; k += 16) {
11892 for (uint32_t n = 1; n <= 8; n++) {
11893 for (uint32_t m = 1; m <= 2; m++) {
11894 GemmMicrokernelTester()
11895 .mr(2)
11896 .nr(8)
11897 .kr(4)
11898 .sr(1)
11899 .m(m)
11900 .n(n)
11901 .k(k)
11902 .iterations(1)
11903 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11904 }
11905 }
11906 }
11907 }
11908
11909 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_gt_8) {
11910 TEST_REQUIRES_ARM_NEON;
11911 for (uint32_t n = 9; n < 16; n++) {
11912 for (size_t k = 1; k <= 80; k += 17) {
11913 GemmMicrokernelTester()
11914 .mr(2)
11915 .nr(8)
11916 .kr(4)
11917 .sr(1)
11918 .m(2)
11919 .n(n)
11920 .k(k)
11921 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11922 }
11923 }
11924 }
11925
11926 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_gt_8_strided_cn) {
11927 TEST_REQUIRES_ARM_NEON;
11928 for (uint32_t n = 9; n < 16; n++) {
11929 for (size_t k = 1; k <= 80; k += 17) {
11930 GemmMicrokernelTester()
11931 .mr(2)
11932 .nr(8)
11933 .kr(4)
11934 .sr(1)
11935 .m(2)
11936 .n(n)
11937 .k(k)
11938 .cn_stride(11)
11939 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11940 }
11941 }
11942 }
11943
11944 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_gt_8_subtile) {
11945 TEST_REQUIRES_ARM_NEON;
11946 for (uint32_t n = 9; n < 16; n++) {
11947 for (size_t k = 1; k <= 80; k += 17) {
11948 for (uint32_t m = 1; m <= 2; m++) {
11949 GemmMicrokernelTester()
11950 .mr(2)
11951 .nr(8)
11952 .kr(4)
11953 .sr(1)
11954 .m(m)
11955 .n(n)
11956 .k(k)
11957 .iterations(1)
11958 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11959 }
11960 }
11961 }
11962 }
11963
11964 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_div_8) {
11965 TEST_REQUIRES_ARM_NEON;
11966 for (uint32_t n = 16; n <= 24; n += 8) {
11967 for (size_t k = 1; k <= 80; k += 17) {
11968 GemmMicrokernelTester()
11969 .mr(2)
11970 .nr(8)
11971 .kr(4)
11972 .sr(1)
11973 .m(2)
11974 .n(n)
11975 .k(k)
11976 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11977 }
11978 }
11979 }
11980
11981 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_div_8_strided_cn) {
11982 TEST_REQUIRES_ARM_NEON;
11983 for (uint32_t n = 16; n <= 24; n += 8) {
11984 for (size_t k = 1; k <= 80; k += 17) {
11985 GemmMicrokernelTester()
11986 .mr(2)
11987 .nr(8)
11988 .kr(4)
11989 .sr(1)
11990 .m(2)
11991 .n(n)
11992 .k(k)
11993 .cn_stride(11)
11994 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
11995 }
11996 }
11997 }
11998
11999 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_div_8_subtile) {
12000 TEST_REQUIRES_ARM_NEON;
12001 for (uint32_t n = 16; n <= 24; n += 8) {
12002 for (size_t k = 1; k <= 80; k += 17) {
12003 for (uint32_t m = 1; m <= 2; m++) {
12004 GemmMicrokernelTester()
12005 .mr(2)
12006 .nr(8)
12007 .kr(4)
12008 .sr(1)
12009 .m(m)
12010 .n(n)
12011 .k(k)
12012 .iterations(1)
12013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12014 }
12015 }
12016 }
12017 }
12018
12019 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, small_kernel) {
12020 TEST_REQUIRES_ARM_NEON;
12021 for (size_t k = 1; k <= 80; k += 17) {
12022 GemmMicrokernelTester()
12023 .mr(2)
12024 .nr(8)
12025 .kr(4)
12026 .sr(1)
12027 .m(2)
12028 .n(8)
12029 .k(k)
12030 .ks(3)
12031 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12032 }
12033 }
12034
12035 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, small_kernel_subtile) {
12036 TEST_REQUIRES_ARM_NEON;
12037 for (size_t k = 1; k <= 80; k += 17) {
12038 for (uint32_t n = 1; n <= 8; n++) {
12039 for (uint32_t m = 1; m <= 2; m++) {
12040 GemmMicrokernelTester()
12041 .mr(2)
12042 .nr(8)
12043 .kr(4)
12044 .sr(1)
12045 .m(m)
12046 .n(n)
12047 .k(k)
12048 .ks(3)
12049 .iterations(1)
12050 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12051 }
12052 }
12053 }
12054 }
12055
12056 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_gt_8_small_kernel) {
12057 TEST_REQUIRES_ARM_NEON;
12058 for (uint32_t n = 9; n < 16; n++) {
12059 for (size_t k = 1; k <= 80; k += 17) {
12060 GemmMicrokernelTester()
12061 .mr(2)
12062 .nr(8)
12063 .kr(4)
12064 .sr(1)
12065 .m(2)
12066 .n(n)
12067 .k(k)
12068 .ks(3)
12069 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12070 }
12071 }
12072 }
12073
12074 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, n_div_8_small_kernel) {
12075 TEST_REQUIRES_ARM_NEON;
12076 for (uint32_t n = 16; n <= 24; n += 8) {
12077 for (size_t k = 1; k <= 80; k += 17) {
12078 GemmMicrokernelTester()
12079 .mr(2)
12080 .nr(8)
12081 .kr(4)
12082 .sr(1)
12083 .m(2)
12084 .n(n)
12085 .k(k)
12086 .ks(3)
12087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12088 }
12089 }
12090 }
12091
12092 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, strided_cm_subtile) {
12093 TEST_REQUIRES_ARM_NEON;
12094 for (size_t k = 1; k <= 80; k += 17) {
12095 for (uint32_t n = 1; n <= 8; n++) {
12096 for (uint32_t m = 1; m <= 2; m++) {
12097 GemmMicrokernelTester()
12098 .mr(2)
12099 .nr(8)
12100 .kr(4)
12101 .sr(1)
12102 .m(m)
12103 .n(n)
12104 .k(k)
12105 .cm_stride(11)
12106 .iterations(1)
12107 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12108 }
12109 }
12110 }
12111 }
12112
12113 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, a_offset) {
12114 TEST_REQUIRES_ARM_NEON;
12115 for (size_t k = 1; k <= 80; k += 17) {
12116 GemmMicrokernelTester()
12117 .mr(2)
12118 .nr(8)
12119 .kr(4)
12120 .sr(1)
12121 .m(2)
12122 .n(8)
12123 .k(k)
12124 .ks(3)
12125 .a_offset(163)
12126 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12127 }
12128 }
12129
12130 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, zero) {
12131 TEST_REQUIRES_ARM_NEON;
12132 for (size_t k = 1; k <= 80; k += 17) {
12133 for (uint32_t mz = 0; mz < 2; mz++) {
12134 GemmMicrokernelTester()
12135 .mr(2)
12136 .nr(8)
12137 .kr(4)
12138 .sr(1)
12139 .m(2)
12140 .n(8)
12141 .k(k)
12142 .ks(3)
12143 .a_offset(163)
12144 .zero_index(mz)
12145 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12146 }
12147 }
12148 }
12149
12150 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, qmin) {
12151 TEST_REQUIRES_ARM_NEON;
12152 GemmMicrokernelTester()
12153 .mr(2)
12154 .nr(8)
12155 .kr(4)
12156 .sr(1)
12157 .m(2)
12158 .n(8)
12159 .k(16)
12160 .qmin(128)
12161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12162 }
12163
12164 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, qmax) {
12165 TEST_REQUIRES_ARM_NEON;
12166 GemmMicrokernelTester()
12167 .mr(2)
12168 .nr(8)
12169 .kr(4)
12170 .sr(1)
12171 .m(2)
12172 .n(8)
12173 .k(16)
12174 .qmax(128)
12175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12176 }
12177
12178 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C4__NEON_MLAL_LD2R, strided_cm) {
12179 TEST_REQUIRES_ARM_NEON;
12180 GemmMicrokernelTester()
12181 .mr(2)
12182 .nr(8)
12183 .kr(4)
12184 .sr(1)
12185 .m(2)
12186 .n(8)
12187 .k(16)
12188 .cm_stride(11)
12189 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c4__neon_mlal_ld2r, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12190 }
12191#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
12192
12193
12194#if XNN_ARCH_ARM || XNN_ARCH_ARM64
12195 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8) {
12196 TEST_REQUIRES_ARM_NEON;
12197 GemmMicrokernelTester()
12198 .mr(4)
12199 .nr(8)
12200 .kr(2)
12201 .sr(1)
12202 .m(4)
12203 .n(8)
12204 .k(8)
12205 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12206 }
12207
12208 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cn) {
12209 TEST_REQUIRES_ARM_NEON;
12210 GemmMicrokernelTester()
12211 .mr(4)
12212 .nr(8)
12213 .kr(2)
12214 .sr(1)
12215 .m(4)
12216 .n(8)
12217 .k(8)
12218 .cn_stride(11)
12219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12220 }
12221
12222 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile) {
12223 TEST_REQUIRES_ARM_NEON;
12224 for (uint32_t n = 1; n <= 8; n++) {
12225 for (uint32_t m = 1; m <= 4; m++) {
12226 GemmMicrokernelTester()
12227 .mr(4)
12228 .nr(8)
12229 .kr(2)
12230 .sr(1)
12231 .m(m)
12232 .n(n)
12233 .k(8)
12234 .iterations(1)
12235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12236 }
12237 }
12238 }
12239
12240 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile_m) {
12241 TEST_REQUIRES_ARM_NEON;
12242 for (uint32_t m = 1; m <= 4; m++) {
12243 GemmMicrokernelTester()
12244 .mr(4)
12245 .nr(8)
12246 .kr(2)
12247 .sr(1)
12248 .m(m)
12249 .n(8)
12250 .k(8)
12251 .iterations(1)
12252 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12253 }
12254 }
12255
12256 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_eq_8_subtile_n) {
12257 TEST_REQUIRES_ARM_NEON;
12258 for (uint32_t n = 1; n <= 8; n++) {
12259 GemmMicrokernelTester()
12260 .mr(4)
12261 .nr(8)
12262 .kr(2)
12263 .sr(1)
12264 .m(4)
12265 .n(n)
12266 .k(8)
12267 .iterations(1)
12268 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12269 }
12270 }
12271
12272 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_lt_8) {
12273 TEST_REQUIRES_ARM_NEON;
12274 for (size_t k = 1; k < 8; k++) {
12275 GemmMicrokernelTester()
12276 .mr(4)
12277 .nr(8)
12278 .kr(2)
12279 .sr(1)
12280 .m(4)
12281 .n(8)
12282 .k(k)
12283 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12284 }
12285 }
12286
12287 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_lt_8_subtile) {
12288 TEST_REQUIRES_ARM_NEON;
12289 for (size_t k = 1; k < 8; k++) {
12290 for (uint32_t n = 1; n <= 8; n++) {
12291 for (uint32_t m = 1; m <= 4; m++) {
12292 GemmMicrokernelTester()
12293 .mr(4)
12294 .nr(8)
12295 .kr(2)
12296 .sr(1)
12297 .m(m)
12298 .n(n)
12299 .k(k)
12300 .iterations(1)
12301 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12302 }
12303 }
12304 }
12305 }
12306
12307 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_gt_8) {
12308 TEST_REQUIRES_ARM_NEON;
12309 for (size_t k = 9; k < 16; k++) {
12310 GemmMicrokernelTester()
12311 .mr(4)
12312 .nr(8)
12313 .kr(2)
12314 .sr(1)
12315 .m(4)
12316 .n(8)
12317 .k(k)
12318 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12319 }
12320 }
12321
12322 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_gt_8_subtile) {
12323 TEST_REQUIRES_ARM_NEON;
12324 for (size_t k = 9; k < 16; k++) {
12325 for (uint32_t n = 1; n <= 8; n++) {
12326 for (uint32_t m = 1; m <= 4; m++) {
12327 GemmMicrokernelTester()
12328 .mr(4)
12329 .nr(8)
12330 .kr(2)
12331 .sr(1)
12332 .m(m)
12333 .n(n)
12334 .k(k)
12335 .iterations(1)
12336 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12337 }
12338 }
12339 }
12340 }
12341
12342 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_div_8) {
12343 TEST_REQUIRES_ARM_NEON;
12344 for (size_t k = 16; k <= 80; k += 8) {
12345 GemmMicrokernelTester()
12346 .mr(4)
12347 .nr(8)
12348 .kr(2)
12349 .sr(1)
12350 .m(4)
12351 .n(8)
12352 .k(k)
12353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12354 }
12355 }
12356
12357 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, k_div_8_subtile) {
12358 TEST_REQUIRES_ARM_NEON;
12359 for (size_t k = 16; k <= 80; k += 8) {
12360 for (uint32_t n = 1; n <= 8; n++) {
12361 for (uint32_t m = 1; m <= 4; m++) {
12362 GemmMicrokernelTester()
12363 .mr(4)
12364 .nr(8)
12365 .kr(2)
12366 .sr(1)
12367 .m(m)
12368 .n(n)
12369 .k(k)
12370 .iterations(1)
12371 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12372 }
12373 }
12374 }
12375 }
12376
12377 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8) {
12378 TEST_REQUIRES_ARM_NEON;
12379 for (uint32_t n = 9; n < 16; n++) {
12380 for (size_t k = 1; k <= 40; k += 9) {
12381 GemmMicrokernelTester()
12382 .mr(4)
12383 .nr(8)
12384 .kr(2)
12385 .sr(1)
12386 .m(4)
12387 .n(n)
12388 .k(k)
12389 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12390 }
12391 }
12392 }
12393
12394 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_strided_cn) {
12395 TEST_REQUIRES_ARM_NEON;
12396 for (uint32_t n = 9; n < 16; n++) {
12397 for (size_t k = 1; k <= 40; k += 9) {
12398 GemmMicrokernelTester()
12399 .mr(4)
12400 .nr(8)
12401 .kr(2)
12402 .sr(1)
12403 .m(4)
12404 .n(n)
12405 .k(k)
12406 .cn_stride(11)
12407 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12408 }
12409 }
12410 }
12411
12412 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_subtile) {
12413 TEST_REQUIRES_ARM_NEON;
12414 for (uint32_t n = 9; n < 16; n++) {
12415 for (size_t k = 1; k <= 40; k += 9) {
12416 for (uint32_t m = 1; m <= 4; m++) {
12417 GemmMicrokernelTester()
12418 .mr(4)
12419 .nr(8)
12420 .kr(2)
12421 .sr(1)
12422 .m(m)
12423 .n(n)
12424 .k(k)
12425 .iterations(1)
12426 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12427 }
12428 }
12429 }
12430 }
12431
12432 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8) {
12433 TEST_REQUIRES_ARM_NEON;
12434 for (uint32_t n = 16; n <= 24; n += 8) {
12435 for (size_t k = 1; k <= 40; k += 9) {
12436 GemmMicrokernelTester()
12437 .mr(4)
12438 .nr(8)
12439 .kr(2)
12440 .sr(1)
12441 .m(4)
12442 .n(n)
12443 .k(k)
12444 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12445 }
12446 }
12447 }
12448
12449 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_strided_cn) {
12450 TEST_REQUIRES_ARM_NEON;
12451 for (uint32_t n = 16; n <= 24; n += 8) {
12452 for (size_t k = 1; k <= 40; k += 9) {
12453 GemmMicrokernelTester()
12454 .mr(4)
12455 .nr(8)
12456 .kr(2)
12457 .sr(1)
12458 .m(4)
12459 .n(n)
12460 .k(k)
12461 .cn_stride(11)
12462 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12463 }
12464 }
12465 }
12466
12467 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_subtile) {
12468 TEST_REQUIRES_ARM_NEON;
12469 for (uint32_t n = 16; n <= 24; n += 8) {
12470 for (size_t k = 1; k <= 40; k += 9) {
12471 for (uint32_t m = 1; m <= 4; m++) {
12472 GemmMicrokernelTester()
12473 .mr(4)
12474 .nr(8)
12475 .kr(2)
12476 .sr(1)
12477 .m(m)
12478 .n(n)
12479 .k(k)
12480 .iterations(1)
12481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12482 }
12483 }
12484 }
12485 }
12486
12487 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, small_kernel) {
12488 TEST_REQUIRES_ARM_NEON;
12489 for (size_t k = 1; k <= 40; k += 9) {
12490 GemmMicrokernelTester()
12491 .mr(4)
12492 .nr(8)
12493 .kr(2)
12494 .sr(1)
12495 .m(4)
12496 .n(8)
12497 .k(k)
12498 .ks(3)
12499 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12500 }
12501 }
12502
12503 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, small_kernel_subtile) {
12504 TEST_REQUIRES_ARM_NEON;
12505 for (size_t k = 1; k <= 40; k += 9) {
12506 for (uint32_t n = 1; n <= 8; n++) {
12507 for (uint32_t m = 1; m <= 4; m++) {
12508 GemmMicrokernelTester()
12509 .mr(4)
12510 .nr(8)
12511 .kr(2)
12512 .sr(1)
12513 .m(m)
12514 .n(n)
12515 .k(k)
12516 .ks(3)
12517 .iterations(1)
12518 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12519 }
12520 }
12521 }
12522 }
12523
12524 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_gt_8_small_kernel) {
12525 TEST_REQUIRES_ARM_NEON;
12526 for (uint32_t n = 9; n < 16; n++) {
12527 for (size_t k = 1; k <= 40; k += 9) {
12528 GemmMicrokernelTester()
12529 .mr(4)
12530 .nr(8)
12531 .kr(2)
12532 .sr(1)
12533 .m(4)
12534 .n(n)
12535 .k(k)
12536 .ks(3)
12537 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12538 }
12539 }
12540 }
12541
12542 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, n_div_8_small_kernel) {
12543 TEST_REQUIRES_ARM_NEON;
12544 for (uint32_t n = 16; n <= 24; n += 8) {
12545 for (size_t k = 1; k <= 40; k += 9) {
12546 GemmMicrokernelTester()
12547 .mr(4)
12548 .nr(8)
12549 .kr(2)
12550 .sr(1)
12551 .m(4)
12552 .n(n)
12553 .k(k)
12554 .ks(3)
12555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12556 }
12557 }
12558 }
12559
12560 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cm_subtile) {
12561 TEST_REQUIRES_ARM_NEON;
12562 for (size_t k = 1; k <= 40; k += 9) {
12563 for (uint32_t n = 1; n <= 8; n++) {
12564 for (uint32_t m = 1; m <= 4; m++) {
12565 GemmMicrokernelTester()
12566 .mr(4)
12567 .nr(8)
12568 .kr(2)
12569 .sr(1)
12570 .m(m)
12571 .n(n)
12572 .k(k)
12573 .cm_stride(11)
12574 .iterations(1)
12575 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12576 }
12577 }
12578 }
12579 }
12580
12581 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, a_offset) {
12582 TEST_REQUIRES_ARM_NEON;
12583 for (size_t k = 1; k <= 40; k += 9) {
12584 GemmMicrokernelTester()
12585 .mr(4)
12586 .nr(8)
12587 .kr(2)
12588 .sr(1)
12589 .m(4)
12590 .n(8)
12591 .k(k)
12592 .ks(3)
12593 .a_offset(163)
12594 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12595 }
12596 }
12597
12598 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, zero) {
12599 TEST_REQUIRES_ARM_NEON;
12600 for (size_t k = 1; k <= 40; k += 9) {
12601 for (uint32_t mz = 0; mz < 4; mz++) {
12602 GemmMicrokernelTester()
12603 .mr(4)
12604 .nr(8)
12605 .kr(2)
12606 .sr(1)
12607 .m(4)
12608 .n(8)
12609 .k(k)
12610 .ks(3)
12611 .a_offset(163)
12612 .zero_index(mz)
12613 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12614 }
12615 }
12616 }
12617
12618 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, qmin) {
12619 TEST_REQUIRES_ARM_NEON;
12620 GemmMicrokernelTester()
12621 .mr(4)
12622 .nr(8)
12623 .kr(2)
12624 .sr(1)
12625 .m(4)
12626 .n(8)
12627 .k(8)
12628 .qmin(128)
12629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12630 }
12631
12632 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, qmax) {
12633 TEST_REQUIRES_ARM_NEON;
12634 GemmMicrokernelTester()
12635 .mr(4)
12636 .nr(8)
12637 .kr(2)
12638 .sr(1)
12639 .m(4)
12640 .n(8)
12641 .k(8)
12642 .qmax(128)
12643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12644 }
12645
12646 TEST(QS8_IGEMM_MINMAX_RNDNU_4X8C2__NEON_MULL_DUP, strided_cm) {
12647 TEST_REQUIRES_ARM_NEON;
12648 GemmMicrokernelTester()
12649 .mr(4)
12650 .nr(8)
12651 .kr(2)
12652 .sr(1)
12653 .m(4)
12654 .n(8)
12655 .k(8)
12656 .cm_stride(11)
12657 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x8c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12658 }
12659#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
12660
12661
12662#if XNN_ARCH_ARM || XNN_ARCH_ARM64
12663 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_eq_8) {
12664 TEST_REQUIRES_ARM_NEON;
12665 GemmMicrokernelTester()
12666 .mr(2)
12667 .nr(16)
12668 .kr(2)
12669 .sr(1)
12670 .m(2)
12671 .n(16)
12672 .k(8)
12673 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12674 }
12675
12676 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, strided_cn) {
12677 TEST_REQUIRES_ARM_NEON;
12678 GemmMicrokernelTester()
12679 .mr(2)
12680 .nr(16)
12681 .kr(2)
12682 .sr(1)
12683 .m(2)
12684 .n(16)
12685 .k(8)
12686 .cn_stride(19)
12687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12688 }
12689
12690 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_eq_8_subtile) {
12691 TEST_REQUIRES_ARM_NEON;
12692 for (uint32_t n = 1; n <= 16; n++) {
12693 for (uint32_t m = 1; m <= 2; m++) {
12694 GemmMicrokernelTester()
12695 .mr(2)
12696 .nr(16)
12697 .kr(2)
12698 .sr(1)
12699 .m(m)
12700 .n(n)
12701 .k(8)
12702 .iterations(1)
12703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12704 }
12705 }
12706 }
12707
12708 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_eq_8_subtile_m) {
12709 TEST_REQUIRES_ARM_NEON;
12710 for (uint32_t m = 1; m <= 2; m++) {
12711 GemmMicrokernelTester()
12712 .mr(2)
12713 .nr(16)
12714 .kr(2)
12715 .sr(1)
12716 .m(m)
12717 .n(16)
12718 .k(8)
12719 .iterations(1)
12720 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12721 }
12722 }
12723
12724 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_eq_8_subtile_n) {
12725 TEST_REQUIRES_ARM_NEON;
12726 for (uint32_t n = 1; n <= 16; n++) {
12727 GemmMicrokernelTester()
12728 .mr(2)
12729 .nr(16)
12730 .kr(2)
12731 .sr(1)
12732 .m(2)
12733 .n(n)
12734 .k(8)
12735 .iterations(1)
12736 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12737 }
12738 }
12739
12740 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_lt_8) {
12741 TEST_REQUIRES_ARM_NEON;
12742 for (size_t k = 1; k < 8; k++) {
12743 GemmMicrokernelTester()
12744 .mr(2)
12745 .nr(16)
12746 .kr(2)
12747 .sr(1)
12748 .m(2)
12749 .n(16)
12750 .k(k)
12751 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12752 }
12753 }
12754
12755 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_lt_8_subtile) {
12756 TEST_REQUIRES_ARM_NEON;
12757 for (size_t k = 1; k < 8; k++) {
12758 for (uint32_t n = 1; n <= 16; n++) {
12759 for (uint32_t m = 1; m <= 2; m++) {
12760 GemmMicrokernelTester()
12761 .mr(2)
12762 .nr(16)
12763 .kr(2)
12764 .sr(1)
12765 .m(m)
12766 .n(n)
12767 .k(k)
12768 .iterations(1)
12769 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12770 }
12771 }
12772 }
12773 }
12774
12775 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_gt_8) {
12776 TEST_REQUIRES_ARM_NEON;
12777 for (size_t k = 9; k < 16; k++) {
12778 GemmMicrokernelTester()
12779 .mr(2)
12780 .nr(16)
12781 .kr(2)
12782 .sr(1)
12783 .m(2)
12784 .n(16)
12785 .k(k)
12786 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12787 }
12788 }
12789
12790 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_gt_8_subtile) {
12791 TEST_REQUIRES_ARM_NEON;
12792 for (size_t k = 9; k < 16; k++) {
12793 for (uint32_t n = 1; n <= 16; n++) {
12794 for (uint32_t m = 1; m <= 2; m++) {
12795 GemmMicrokernelTester()
12796 .mr(2)
12797 .nr(16)
12798 .kr(2)
12799 .sr(1)
12800 .m(m)
12801 .n(n)
12802 .k(k)
12803 .iterations(1)
12804 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12805 }
12806 }
12807 }
12808 }
12809
12810 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_div_8) {
12811 TEST_REQUIRES_ARM_NEON;
12812 for (size_t k = 16; k <= 80; k += 8) {
12813 GemmMicrokernelTester()
12814 .mr(2)
12815 .nr(16)
12816 .kr(2)
12817 .sr(1)
12818 .m(2)
12819 .n(16)
12820 .k(k)
12821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12822 }
12823 }
12824
12825 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, k_div_8_subtile) {
12826 TEST_REQUIRES_ARM_NEON;
12827 for (size_t k = 16; k <= 80; k += 8) {
12828 for (uint32_t n = 1; n <= 16; n++) {
12829 for (uint32_t m = 1; m <= 2; m++) {
12830 GemmMicrokernelTester()
12831 .mr(2)
12832 .nr(16)
12833 .kr(2)
12834 .sr(1)
12835 .m(m)
12836 .n(n)
12837 .k(k)
12838 .iterations(1)
12839 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12840 }
12841 }
12842 }
12843 }
12844
12845 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_gt_16) {
12846 TEST_REQUIRES_ARM_NEON;
12847 for (uint32_t n = 17; n < 32; n++) {
12848 for (size_t k = 1; k <= 40; k += 9) {
12849 GemmMicrokernelTester()
12850 .mr(2)
12851 .nr(16)
12852 .kr(2)
12853 .sr(1)
12854 .m(2)
12855 .n(n)
12856 .k(k)
12857 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12858 }
12859 }
12860 }
12861
12862 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_gt_16_strided_cn) {
12863 TEST_REQUIRES_ARM_NEON;
12864 for (uint32_t n = 17; n < 32; n++) {
12865 for (size_t k = 1; k <= 40; k += 9) {
12866 GemmMicrokernelTester()
12867 .mr(2)
12868 .nr(16)
12869 .kr(2)
12870 .sr(1)
12871 .m(2)
12872 .n(n)
12873 .k(k)
12874 .cn_stride(19)
12875 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12876 }
12877 }
12878 }
12879
12880 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_gt_16_subtile) {
12881 TEST_REQUIRES_ARM_NEON;
12882 for (uint32_t n = 17; n < 32; n++) {
12883 for (size_t k = 1; k <= 40; k += 9) {
12884 for (uint32_t m = 1; m <= 2; m++) {
12885 GemmMicrokernelTester()
12886 .mr(2)
12887 .nr(16)
12888 .kr(2)
12889 .sr(1)
12890 .m(m)
12891 .n(n)
12892 .k(k)
12893 .iterations(1)
12894 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12895 }
12896 }
12897 }
12898 }
12899
12900 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_div_16) {
12901 TEST_REQUIRES_ARM_NEON;
12902 for (uint32_t n = 32; n <= 48; n += 16) {
12903 for (size_t k = 1; k <= 40; k += 9) {
12904 GemmMicrokernelTester()
12905 .mr(2)
12906 .nr(16)
12907 .kr(2)
12908 .sr(1)
12909 .m(2)
12910 .n(n)
12911 .k(k)
12912 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12913 }
12914 }
12915 }
12916
12917 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_div_16_strided_cn) {
12918 TEST_REQUIRES_ARM_NEON;
12919 for (uint32_t n = 32; n <= 48; n += 16) {
12920 for (size_t k = 1; k <= 40; k += 9) {
12921 GemmMicrokernelTester()
12922 .mr(2)
12923 .nr(16)
12924 .kr(2)
12925 .sr(1)
12926 .m(2)
12927 .n(n)
12928 .k(k)
12929 .cn_stride(19)
12930 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12931 }
12932 }
12933 }
12934
12935 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_div_16_subtile) {
12936 TEST_REQUIRES_ARM_NEON;
12937 for (uint32_t n = 32; n <= 48; n += 16) {
12938 for (size_t k = 1; k <= 40; k += 9) {
12939 for (uint32_t m = 1; m <= 2; m++) {
12940 GemmMicrokernelTester()
12941 .mr(2)
12942 .nr(16)
12943 .kr(2)
12944 .sr(1)
12945 .m(m)
12946 .n(n)
12947 .k(k)
12948 .iterations(1)
12949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12950 }
12951 }
12952 }
12953 }
12954
12955 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, small_kernel) {
12956 TEST_REQUIRES_ARM_NEON;
12957 for (size_t k = 1; k <= 40; k += 9) {
12958 GemmMicrokernelTester()
12959 .mr(2)
12960 .nr(16)
12961 .kr(2)
12962 .sr(1)
12963 .m(2)
12964 .n(16)
12965 .k(k)
12966 .ks(3)
12967 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12968 }
12969 }
12970
12971 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, small_kernel_subtile) {
12972 TEST_REQUIRES_ARM_NEON;
12973 for (size_t k = 1; k <= 40; k += 9) {
12974 for (uint32_t n = 1; n <= 16; n++) {
12975 for (uint32_t m = 1; m <= 2; m++) {
12976 GemmMicrokernelTester()
12977 .mr(2)
12978 .nr(16)
12979 .kr(2)
12980 .sr(1)
12981 .m(m)
12982 .n(n)
12983 .k(k)
12984 .ks(3)
12985 .iterations(1)
12986 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
12987 }
12988 }
12989 }
12990 }
12991
12992 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_gt_16_small_kernel) {
12993 TEST_REQUIRES_ARM_NEON;
12994 for (uint32_t n = 17; n < 32; n++) {
12995 for (size_t k = 1; k <= 40; k += 9) {
12996 GemmMicrokernelTester()
12997 .mr(2)
12998 .nr(16)
12999 .kr(2)
13000 .sr(1)
13001 .m(2)
13002 .n(n)
13003 .k(k)
13004 .ks(3)
13005 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13006 }
13007 }
13008 }
13009
13010 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, n_div_16_small_kernel) {
13011 TEST_REQUIRES_ARM_NEON;
13012 for (uint32_t n = 32; n <= 48; n += 16) {
13013 for (size_t k = 1; k <= 40; k += 9) {
13014 GemmMicrokernelTester()
13015 .mr(2)
13016 .nr(16)
13017 .kr(2)
13018 .sr(1)
13019 .m(2)
13020 .n(n)
13021 .k(k)
13022 .ks(3)
13023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13024 }
13025 }
13026 }
13027
13028 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, strided_cm_subtile) {
13029 TEST_REQUIRES_ARM_NEON;
13030 for (size_t k = 1; k <= 40; k += 9) {
13031 for (uint32_t n = 1; n <= 16; n++) {
13032 for (uint32_t m = 1; m <= 2; m++) {
13033 GemmMicrokernelTester()
13034 .mr(2)
13035 .nr(16)
13036 .kr(2)
13037 .sr(1)
13038 .m(m)
13039 .n(n)
13040 .k(k)
13041 .cm_stride(19)
13042 .iterations(1)
13043 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13044 }
13045 }
13046 }
13047 }
13048
13049 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, a_offset) {
13050 TEST_REQUIRES_ARM_NEON;
13051 for (size_t k = 1; k <= 40; k += 9) {
13052 GemmMicrokernelTester()
13053 .mr(2)
13054 .nr(16)
13055 .kr(2)
13056 .sr(1)
13057 .m(2)
13058 .n(16)
13059 .k(k)
13060 .ks(3)
13061 .a_offset(83)
13062 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13063 }
13064 }
13065
13066 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, zero) {
13067 TEST_REQUIRES_ARM_NEON;
13068 for (size_t k = 1; k <= 40; k += 9) {
13069 for (uint32_t mz = 0; mz < 2; mz++) {
13070 GemmMicrokernelTester()
13071 .mr(2)
13072 .nr(16)
13073 .kr(2)
13074 .sr(1)
13075 .m(2)
13076 .n(16)
13077 .k(k)
13078 .ks(3)
13079 .a_offset(83)
13080 .zero_index(mz)
13081 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13082 }
13083 }
13084 }
13085
13086 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, qmin) {
13087 TEST_REQUIRES_ARM_NEON;
13088 GemmMicrokernelTester()
13089 .mr(2)
13090 .nr(16)
13091 .kr(2)
13092 .sr(1)
13093 .m(2)
13094 .n(16)
13095 .k(8)
13096 .qmin(128)
13097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13098 }
13099
13100 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, qmax) {
13101 TEST_REQUIRES_ARM_NEON;
13102 GemmMicrokernelTester()
13103 .mr(2)
13104 .nr(16)
13105 .kr(2)
13106 .sr(1)
13107 .m(2)
13108 .n(16)
13109 .k(8)
13110 .qmax(128)
13111 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13112 }
13113
13114 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C2__NEON_MULL_DUP, strided_cm) {
13115 TEST_REQUIRES_ARM_NEON;
13116 GemmMicrokernelTester()
13117 .mr(2)
13118 .nr(16)
13119 .kr(2)
13120 .sr(1)
13121 .m(2)
13122 .n(16)
13123 .k(8)
13124 .cm_stride(19)
13125 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c2__neon_mull_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13126 }
13127#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
13128
13129
13130#if XNN_ARCH_ARM || XNN_ARCH_ARM64
13131 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16) {
13132 TEST_REQUIRES_ARM_NEON;
13133 GemmMicrokernelTester()
13134 .mr(1)
13135 .nr(8)
13136 .kr(2)
13137 .sr(1)
13138 .m(1)
13139 .n(8)
13140 .k(16)
13141 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13142 }
13143
13144 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, strided_cn) {
13145 TEST_REQUIRES_ARM_NEON;
13146 GemmMicrokernelTester()
13147 .mr(1)
13148 .nr(8)
13149 .kr(2)
13150 .sr(1)
13151 .m(1)
13152 .n(8)
13153 .k(16)
13154 .cn_stride(11)
13155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13156 }
13157
13158 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16_subtile) {
13159 TEST_REQUIRES_ARM_NEON;
13160 for (uint32_t n = 1; n <= 8; n++) {
13161 for (uint32_t m = 1; m <= 1; m++) {
13162 GemmMicrokernelTester()
13163 .mr(1)
13164 .nr(8)
13165 .kr(2)
13166 .sr(1)
13167 .m(m)
13168 .n(n)
13169 .k(16)
13170 .iterations(1)
13171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13172 }
13173 }
13174 }
13175
13176 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) {
13177 TEST_REQUIRES_ARM_NEON;
13178 for (uint32_t m = 1; m <= 1; m++) {
13179 GemmMicrokernelTester()
13180 .mr(1)
13181 .nr(8)
13182 .kr(2)
13183 .sr(1)
13184 .m(m)
13185 .n(8)
13186 .k(16)
13187 .iterations(1)
13188 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13189 }
13190 }
13191
13192 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) {
13193 TEST_REQUIRES_ARM_NEON;
13194 for (uint32_t n = 1; n <= 8; n++) {
13195 GemmMicrokernelTester()
13196 .mr(1)
13197 .nr(8)
13198 .kr(2)
13199 .sr(1)
13200 .m(1)
13201 .n(n)
13202 .k(16)
13203 .iterations(1)
13204 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13205 }
13206 }
13207
13208 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_lt_16) {
13209 TEST_REQUIRES_ARM_NEON;
13210 for (size_t k = 1; k < 16; k++) {
13211 GemmMicrokernelTester()
13212 .mr(1)
13213 .nr(8)
13214 .kr(2)
13215 .sr(1)
13216 .m(1)
13217 .n(8)
13218 .k(k)
13219 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13220 }
13221 }
13222
13223 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_lt_16_subtile) {
13224 TEST_REQUIRES_ARM_NEON;
13225 for (size_t k = 1; k < 16; k++) {
13226 for (uint32_t n = 1; n <= 8; n++) {
13227 for (uint32_t m = 1; m <= 1; m++) {
13228 GemmMicrokernelTester()
13229 .mr(1)
13230 .nr(8)
13231 .kr(2)
13232 .sr(1)
13233 .m(m)
13234 .n(n)
13235 .k(k)
13236 .iterations(1)
13237 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13238 }
13239 }
13240 }
13241 }
13242
13243 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_gt_16) {
13244 TEST_REQUIRES_ARM_NEON;
13245 for (size_t k = 17; k < 32; k++) {
13246 GemmMicrokernelTester()
13247 .mr(1)
13248 .nr(8)
13249 .kr(2)
13250 .sr(1)
13251 .m(1)
13252 .n(8)
13253 .k(k)
13254 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13255 }
13256 }
13257
13258 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_gt_16_subtile) {
13259 TEST_REQUIRES_ARM_NEON;
13260 for (size_t k = 17; k < 32; k++) {
13261 for (uint32_t n = 1; n <= 8; n++) {
13262 for (uint32_t m = 1; m <= 1; m++) {
13263 GemmMicrokernelTester()
13264 .mr(1)
13265 .nr(8)
13266 .kr(2)
13267 .sr(1)
13268 .m(m)
13269 .n(n)
13270 .k(k)
13271 .iterations(1)
13272 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13273 }
13274 }
13275 }
13276 }
13277
13278 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_div_16) {
13279 TEST_REQUIRES_ARM_NEON;
13280 for (size_t k = 32; k <= 160; k += 16) {
13281 GemmMicrokernelTester()
13282 .mr(1)
13283 .nr(8)
13284 .kr(2)
13285 .sr(1)
13286 .m(1)
13287 .n(8)
13288 .k(k)
13289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13290 }
13291 }
13292
13293 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, k_div_16_subtile) {
13294 TEST_REQUIRES_ARM_NEON;
13295 for (size_t k = 32; k <= 160; k += 16) {
13296 for (uint32_t n = 1; n <= 8; n++) {
13297 for (uint32_t m = 1; m <= 1; m++) {
13298 GemmMicrokernelTester()
13299 .mr(1)
13300 .nr(8)
13301 .kr(2)
13302 .sr(1)
13303 .m(m)
13304 .n(n)
13305 .k(k)
13306 .iterations(1)
13307 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13308 }
13309 }
13310 }
13311 }
13312
13313 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8) {
13314 TEST_REQUIRES_ARM_NEON;
13315 for (uint32_t n = 9; n < 16; n++) {
13316 for (size_t k = 1; k <= 80; k += 17) {
13317 GemmMicrokernelTester()
13318 .mr(1)
13319 .nr(8)
13320 .kr(2)
13321 .sr(1)
13322 .m(1)
13323 .n(n)
13324 .k(k)
13325 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13326 }
13327 }
13328 }
13329
13330 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) {
13331 TEST_REQUIRES_ARM_NEON;
13332 for (uint32_t n = 9; n < 16; n++) {
13333 for (size_t k = 1; k <= 80; k += 17) {
13334 GemmMicrokernelTester()
13335 .mr(1)
13336 .nr(8)
13337 .kr(2)
13338 .sr(1)
13339 .m(1)
13340 .n(n)
13341 .k(k)
13342 .cn_stride(11)
13343 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13344 }
13345 }
13346 }
13347
13348 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8_subtile) {
13349 TEST_REQUIRES_ARM_NEON;
13350 for (uint32_t n = 9; n < 16; n++) {
13351 for (size_t k = 1; k <= 80; k += 17) {
13352 for (uint32_t m = 1; m <= 1; m++) {
13353 GemmMicrokernelTester()
13354 .mr(1)
13355 .nr(8)
13356 .kr(2)
13357 .sr(1)
13358 .m(m)
13359 .n(n)
13360 .k(k)
13361 .iterations(1)
13362 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13363 }
13364 }
13365 }
13366 }
13367
13368 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8) {
13369 TEST_REQUIRES_ARM_NEON;
13370 for (uint32_t n = 16; n <= 24; n += 8) {
13371 for (size_t k = 1; k <= 80; k += 17) {
13372 GemmMicrokernelTester()
13373 .mr(1)
13374 .nr(8)
13375 .kr(2)
13376 .sr(1)
13377 .m(1)
13378 .n(n)
13379 .k(k)
13380 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13381 }
13382 }
13383 }
13384
13385 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) {
13386 TEST_REQUIRES_ARM_NEON;
13387 for (uint32_t n = 16; n <= 24; n += 8) {
13388 for (size_t k = 1; k <= 80; k += 17) {
13389 GemmMicrokernelTester()
13390 .mr(1)
13391 .nr(8)
13392 .kr(2)
13393 .sr(1)
13394 .m(1)
13395 .n(n)
13396 .k(k)
13397 .cn_stride(11)
13398 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13399 }
13400 }
13401 }
13402
13403 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8_subtile) {
13404 TEST_REQUIRES_ARM_NEON;
13405 for (uint32_t n = 16; n <= 24; n += 8) {
13406 for (size_t k = 1; k <= 80; k += 17) {
13407 for (uint32_t m = 1; m <= 1; m++) {
13408 GemmMicrokernelTester()
13409 .mr(1)
13410 .nr(8)
13411 .kr(2)
13412 .sr(1)
13413 .m(m)
13414 .n(n)
13415 .k(k)
13416 .iterations(1)
13417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13418 }
13419 }
13420 }
13421 }
13422
13423 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, small_kernel) {
13424 TEST_REQUIRES_ARM_NEON;
13425 for (size_t k = 1; k <= 80; k += 17) {
13426 GemmMicrokernelTester()
13427 .mr(1)
13428 .nr(8)
13429 .kr(2)
13430 .sr(1)
13431 .m(1)
13432 .n(8)
13433 .k(k)
13434 .ks(3)
13435 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13436 }
13437 }
13438
13439 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, small_kernel_subtile) {
13440 TEST_REQUIRES_ARM_NEON;
13441 for (size_t k = 1; k <= 80; k += 17) {
13442 for (uint32_t n = 1; n <= 8; n++) {
13443 for (uint32_t m = 1; m <= 1; m++) {
13444 GemmMicrokernelTester()
13445 .mr(1)
13446 .nr(8)
13447 .kr(2)
13448 .sr(1)
13449 .m(m)
13450 .n(n)
13451 .k(k)
13452 .ks(3)
13453 .iterations(1)
13454 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13455 }
13456 }
13457 }
13458 }
13459
13460 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_gt_8_small_kernel) {
13461 TEST_REQUIRES_ARM_NEON;
13462 for (uint32_t n = 9; n < 16; n++) {
13463 for (size_t k = 1; k <= 80; k += 17) {
13464 GemmMicrokernelTester()
13465 .mr(1)
13466 .nr(8)
13467 .kr(2)
13468 .sr(1)
13469 .m(1)
13470 .n(n)
13471 .k(k)
13472 .ks(3)
13473 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13474 }
13475 }
13476 }
13477
13478 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, n_div_8_small_kernel) {
13479 TEST_REQUIRES_ARM_NEON;
13480 for (uint32_t n = 16; n <= 24; n += 8) {
13481 for (size_t k = 1; k <= 80; k += 17) {
13482 GemmMicrokernelTester()
13483 .mr(1)
13484 .nr(8)
13485 .kr(2)
13486 .sr(1)
13487 .m(1)
13488 .n(n)
13489 .k(k)
13490 .ks(3)
13491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13492 }
13493 }
13494 }
13495
13496 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, strided_cm_subtile) {
13497 TEST_REQUIRES_ARM_NEON;
13498 for (size_t k = 1; k <= 80; k += 17) {
13499 for (uint32_t n = 1; n <= 8; n++) {
13500 for (uint32_t m = 1; m <= 1; m++) {
13501 GemmMicrokernelTester()
13502 .mr(1)
13503 .nr(8)
13504 .kr(2)
13505 .sr(1)
13506 .m(m)
13507 .n(n)
13508 .k(k)
13509 .cm_stride(11)
13510 .iterations(1)
13511 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13512 }
13513 }
13514 }
13515 }
13516
13517 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, a_offset) {
13518 TEST_REQUIRES_ARM_NEON;
13519 for (size_t k = 1; k <= 80; k += 17) {
13520 GemmMicrokernelTester()
13521 .mr(1)
13522 .nr(8)
13523 .kr(2)
13524 .sr(1)
13525 .m(1)
13526 .n(8)
13527 .k(k)
13528 .ks(3)
13529 .a_offset(83)
13530 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13531 }
13532 }
13533
13534 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, zero) {
13535 TEST_REQUIRES_ARM_NEON;
13536 for (size_t k = 1; k <= 80; k += 17) {
13537 for (uint32_t mz = 0; mz < 1; mz++) {
13538 GemmMicrokernelTester()
13539 .mr(1)
13540 .nr(8)
13541 .kr(2)
13542 .sr(1)
13543 .m(1)
13544 .n(8)
13545 .k(k)
13546 .ks(3)
13547 .a_offset(83)
13548 .zero_index(mz)
13549 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13550 }
13551 }
13552 }
13553
13554 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, qmin) {
13555 TEST_REQUIRES_ARM_NEON;
13556 GemmMicrokernelTester()
13557 .mr(1)
13558 .nr(8)
13559 .kr(2)
13560 .sr(1)
13561 .m(1)
13562 .n(8)
13563 .k(16)
13564 .qmin(128)
13565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13566 }
13567
13568 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, qmax) {
13569 TEST_REQUIRES_ARM_NEON;
13570 GemmMicrokernelTester()
13571 .mr(1)
13572 .nr(8)
13573 .kr(2)
13574 .sr(1)
13575 .m(1)
13576 .n(8)
13577 .k(16)
13578 .qmax(128)
13579 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13580 }
13581
13582 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C2__NEON_MLAL_DUP, strided_cm) {
13583 TEST_REQUIRES_ARM_NEON;
13584 GemmMicrokernelTester()
13585 .mr(1)
13586 .nr(8)
13587 .kr(2)
13588 .sr(1)
13589 .m(1)
13590 .n(8)
13591 .k(16)
13592 .cm_stride(11)
13593 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13594 }
13595#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
13596
13597
13598#if XNN_ARCH_ARM || XNN_ARCH_ARM64
13599 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16) {
13600 TEST_REQUIRES_ARM_NEON;
13601 GemmMicrokernelTester()
13602 .mr(3)
13603 .nr(8)
13604 .kr(2)
13605 .sr(1)
13606 .m(3)
13607 .n(8)
13608 .k(16)
13609 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13610 }
13611
13612 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cn) {
13613 TEST_REQUIRES_ARM_NEON;
13614 GemmMicrokernelTester()
13615 .mr(3)
13616 .nr(8)
13617 .kr(2)
13618 .sr(1)
13619 .m(3)
13620 .n(8)
13621 .k(16)
13622 .cn_stride(11)
13623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13624 }
13625
13626 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile) {
13627 TEST_REQUIRES_ARM_NEON;
13628 for (uint32_t n = 1; n <= 8; n++) {
13629 for (uint32_t m = 1; m <= 3; m++) {
13630 GemmMicrokernelTester()
13631 .mr(3)
13632 .nr(8)
13633 .kr(2)
13634 .sr(1)
13635 .m(m)
13636 .n(n)
13637 .k(16)
13638 .iterations(1)
13639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13640 }
13641 }
13642 }
13643
13644 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile_m) {
13645 TEST_REQUIRES_ARM_NEON;
13646 for (uint32_t m = 1; m <= 3; m++) {
13647 GemmMicrokernelTester()
13648 .mr(3)
13649 .nr(8)
13650 .kr(2)
13651 .sr(1)
13652 .m(m)
13653 .n(8)
13654 .k(16)
13655 .iterations(1)
13656 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13657 }
13658 }
13659
13660 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_eq_16_subtile_n) {
13661 TEST_REQUIRES_ARM_NEON;
13662 for (uint32_t n = 1; n <= 8; n++) {
13663 GemmMicrokernelTester()
13664 .mr(3)
13665 .nr(8)
13666 .kr(2)
13667 .sr(1)
13668 .m(3)
13669 .n(n)
13670 .k(16)
13671 .iterations(1)
13672 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13673 }
13674 }
13675
13676 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_lt_16) {
13677 TEST_REQUIRES_ARM_NEON;
13678 for (size_t k = 1; k < 16; k++) {
13679 GemmMicrokernelTester()
13680 .mr(3)
13681 .nr(8)
13682 .kr(2)
13683 .sr(1)
13684 .m(3)
13685 .n(8)
13686 .k(k)
13687 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13688 }
13689 }
13690
13691 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_lt_16_subtile) {
13692 TEST_REQUIRES_ARM_NEON;
13693 for (size_t k = 1; k < 16; k++) {
13694 for (uint32_t n = 1; n <= 8; n++) {
13695 for (uint32_t m = 1; m <= 3; m++) {
13696 GemmMicrokernelTester()
13697 .mr(3)
13698 .nr(8)
13699 .kr(2)
13700 .sr(1)
13701 .m(m)
13702 .n(n)
13703 .k(k)
13704 .iterations(1)
13705 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13706 }
13707 }
13708 }
13709 }
13710
13711 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_gt_16) {
13712 TEST_REQUIRES_ARM_NEON;
13713 for (size_t k = 17; k < 32; k++) {
13714 GemmMicrokernelTester()
13715 .mr(3)
13716 .nr(8)
13717 .kr(2)
13718 .sr(1)
13719 .m(3)
13720 .n(8)
13721 .k(k)
13722 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13723 }
13724 }
13725
13726 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_gt_16_subtile) {
13727 TEST_REQUIRES_ARM_NEON;
13728 for (size_t k = 17; k < 32; k++) {
13729 for (uint32_t n = 1; n <= 8; n++) {
13730 for (uint32_t m = 1; m <= 3; m++) {
13731 GemmMicrokernelTester()
13732 .mr(3)
13733 .nr(8)
13734 .kr(2)
13735 .sr(1)
13736 .m(m)
13737 .n(n)
13738 .k(k)
13739 .iterations(1)
13740 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13741 }
13742 }
13743 }
13744 }
13745
13746 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_div_16) {
13747 TEST_REQUIRES_ARM_NEON;
13748 for (size_t k = 32; k <= 160; k += 16) {
13749 GemmMicrokernelTester()
13750 .mr(3)
13751 .nr(8)
13752 .kr(2)
13753 .sr(1)
13754 .m(3)
13755 .n(8)
13756 .k(k)
13757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13758 }
13759 }
13760
13761 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, k_div_16_subtile) {
13762 TEST_REQUIRES_ARM_NEON;
13763 for (size_t k = 32; k <= 160; k += 16) {
13764 for (uint32_t n = 1; n <= 8; n++) {
13765 for (uint32_t m = 1; m <= 3; m++) {
13766 GemmMicrokernelTester()
13767 .mr(3)
13768 .nr(8)
13769 .kr(2)
13770 .sr(1)
13771 .m(m)
13772 .n(n)
13773 .k(k)
13774 .iterations(1)
13775 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13776 }
13777 }
13778 }
13779 }
13780
13781 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8) {
13782 TEST_REQUIRES_ARM_NEON;
13783 for (uint32_t n = 9; n < 16; n++) {
13784 for (size_t k = 1; k <= 80; k += 17) {
13785 GemmMicrokernelTester()
13786 .mr(3)
13787 .nr(8)
13788 .kr(2)
13789 .sr(1)
13790 .m(3)
13791 .n(n)
13792 .k(k)
13793 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13794 }
13795 }
13796 }
13797
13798 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_strided_cn) {
13799 TEST_REQUIRES_ARM_NEON;
13800 for (uint32_t n = 9; n < 16; n++) {
13801 for (size_t k = 1; k <= 80; k += 17) {
13802 GemmMicrokernelTester()
13803 .mr(3)
13804 .nr(8)
13805 .kr(2)
13806 .sr(1)
13807 .m(3)
13808 .n(n)
13809 .k(k)
13810 .cn_stride(11)
13811 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13812 }
13813 }
13814 }
13815
13816 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_subtile) {
13817 TEST_REQUIRES_ARM_NEON;
13818 for (uint32_t n = 9; n < 16; n++) {
13819 for (size_t k = 1; k <= 80; k += 17) {
13820 for (uint32_t m = 1; m <= 3; m++) {
13821 GemmMicrokernelTester()
13822 .mr(3)
13823 .nr(8)
13824 .kr(2)
13825 .sr(1)
13826 .m(m)
13827 .n(n)
13828 .k(k)
13829 .iterations(1)
13830 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13831 }
13832 }
13833 }
13834 }
13835
13836 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8) {
13837 TEST_REQUIRES_ARM_NEON;
13838 for (uint32_t n = 16; n <= 24; n += 8) {
13839 for (size_t k = 1; k <= 80; k += 17) {
13840 GemmMicrokernelTester()
13841 .mr(3)
13842 .nr(8)
13843 .kr(2)
13844 .sr(1)
13845 .m(3)
13846 .n(n)
13847 .k(k)
13848 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13849 }
13850 }
13851 }
13852
13853 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_strided_cn) {
13854 TEST_REQUIRES_ARM_NEON;
13855 for (uint32_t n = 16; n <= 24; n += 8) {
13856 for (size_t k = 1; k <= 80; k += 17) {
13857 GemmMicrokernelTester()
13858 .mr(3)
13859 .nr(8)
13860 .kr(2)
13861 .sr(1)
13862 .m(3)
13863 .n(n)
13864 .k(k)
13865 .cn_stride(11)
13866 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13867 }
13868 }
13869 }
13870
13871 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_subtile) {
13872 TEST_REQUIRES_ARM_NEON;
13873 for (uint32_t n = 16; n <= 24; n += 8) {
13874 for (size_t k = 1; k <= 80; k += 17) {
13875 for (uint32_t m = 1; m <= 3; m++) {
13876 GemmMicrokernelTester()
13877 .mr(3)
13878 .nr(8)
13879 .kr(2)
13880 .sr(1)
13881 .m(m)
13882 .n(n)
13883 .k(k)
13884 .iterations(1)
13885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13886 }
13887 }
13888 }
13889 }
13890
13891 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, small_kernel) {
13892 TEST_REQUIRES_ARM_NEON;
13893 for (size_t k = 1; k <= 80; k += 17) {
13894 GemmMicrokernelTester()
13895 .mr(3)
13896 .nr(8)
13897 .kr(2)
13898 .sr(1)
13899 .m(3)
13900 .n(8)
13901 .k(k)
13902 .ks(3)
13903 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13904 }
13905 }
13906
13907 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, small_kernel_subtile) {
13908 TEST_REQUIRES_ARM_NEON;
13909 for (size_t k = 1; k <= 80; k += 17) {
13910 for (uint32_t n = 1; n <= 8; n++) {
13911 for (uint32_t m = 1; m <= 3; m++) {
13912 GemmMicrokernelTester()
13913 .mr(3)
13914 .nr(8)
13915 .kr(2)
13916 .sr(1)
13917 .m(m)
13918 .n(n)
13919 .k(k)
13920 .ks(3)
13921 .iterations(1)
13922 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13923 }
13924 }
13925 }
13926 }
13927
13928 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_gt_8_small_kernel) {
13929 TEST_REQUIRES_ARM_NEON;
13930 for (uint32_t n = 9; n < 16; n++) {
13931 for (size_t k = 1; k <= 80; k += 17) {
13932 GemmMicrokernelTester()
13933 .mr(3)
13934 .nr(8)
13935 .kr(2)
13936 .sr(1)
13937 .m(3)
13938 .n(n)
13939 .k(k)
13940 .ks(3)
13941 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13942 }
13943 }
13944 }
13945
13946 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, n_div_8_small_kernel) {
13947 TEST_REQUIRES_ARM_NEON;
13948 for (uint32_t n = 16; n <= 24; n += 8) {
13949 for (size_t k = 1; k <= 80; k += 17) {
13950 GemmMicrokernelTester()
13951 .mr(3)
13952 .nr(8)
13953 .kr(2)
13954 .sr(1)
13955 .m(3)
13956 .n(n)
13957 .k(k)
13958 .ks(3)
13959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13960 }
13961 }
13962 }
13963
13964 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cm_subtile) {
13965 TEST_REQUIRES_ARM_NEON;
13966 for (size_t k = 1; k <= 80; k += 17) {
13967 for (uint32_t n = 1; n <= 8; n++) {
13968 for (uint32_t m = 1; m <= 3; m++) {
13969 GemmMicrokernelTester()
13970 .mr(3)
13971 .nr(8)
13972 .kr(2)
13973 .sr(1)
13974 .m(m)
13975 .n(n)
13976 .k(k)
13977 .cm_stride(11)
13978 .iterations(1)
13979 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13980 }
13981 }
13982 }
13983 }
13984
13985 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, a_offset) {
13986 TEST_REQUIRES_ARM_NEON;
13987 for (size_t k = 1; k <= 80; k += 17) {
13988 GemmMicrokernelTester()
13989 .mr(3)
13990 .nr(8)
13991 .kr(2)
13992 .sr(1)
13993 .m(3)
13994 .n(8)
13995 .k(k)
13996 .ks(3)
13997 .a_offset(251)
13998 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
13999 }
14000 }
14001
14002 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, zero) {
14003 TEST_REQUIRES_ARM_NEON;
14004 for (size_t k = 1; k <= 80; k += 17) {
14005 for (uint32_t mz = 0; mz < 3; mz++) {
14006 GemmMicrokernelTester()
14007 .mr(3)
14008 .nr(8)
14009 .kr(2)
14010 .sr(1)
14011 .m(3)
14012 .n(8)
14013 .k(k)
14014 .ks(3)
14015 .a_offset(251)
14016 .zero_index(mz)
14017 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14018 }
14019 }
14020 }
14021
14022 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, qmin) {
14023 TEST_REQUIRES_ARM_NEON;
14024 GemmMicrokernelTester()
14025 .mr(3)
14026 .nr(8)
14027 .kr(2)
14028 .sr(1)
14029 .m(3)
14030 .n(8)
14031 .k(16)
14032 .qmin(128)
14033 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14034 }
14035
14036 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, qmax) {
14037 TEST_REQUIRES_ARM_NEON;
14038 GemmMicrokernelTester()
14039 .mr(3)
14040 .nr(8)
14041 .kr(2)
14042 .sr(1)
14043 .m(3)
14044 .n(8)
14045 .k(16)
14046 .qmax(128)
14047 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14048 }
14049
14050 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C2__NEON_MLAL_DUP, strided_cm) {
14051 TEST_REQUIRES_ARM_NEON;
14052 GemmMicrokernelTester()
14053 .mr(3)
14054 .nr(8)
14055 .kr(2)
14056 .sr(1)
14057 .m(3)
14058 .n(8)
14059 .k(16)
14060 .cm_stride(11)
14061 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14062 }
14063#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
14064
14065
14066#if XNN_ARCH_ARM || XNN_ARCH_ARM64
14067 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16) {
14068 TEST_REQUIRES_ARM_NEON;
14069 GemmMicrokernelTester()
14070 .mr(1)
14071 .nr(16)
14072 .kr(2)
14073 .sr(1)
14074 .m(1)
14075 .n(16)
14076 .k(16)
14077 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14078 }
14079
14080 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, strided_cn) {
14081 TEST_REQUIRES_ARM_NEON;
14082 GemmMicrokernelTester()
14083 .mr(1)
14084 .nr(16)
14085 .kr(2)
14086 .sr(1)
14087 .m(1)
14088 .n(16)
14089 .k(16)
14090 .cn_stride(19)
14091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14092 }
14093
14094 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16_subtile) {
14095 TEST_REQUIRES_ARM_NEON;
14096 for (uint32_t n = 1; n <= 16; n++) {
14097 for (uint32_t m = 1; m <= 1; m++) {
14098 GemmMicrokernelTester()
14099 .mr(1)
14100 .nr(16)
14101 .kr(2)
14102 .sr(1)
14103 .m(m)
14104 .n(n)
14105 .k(16)
14106 .iterations(1)
14107 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14108 }
14109 }
14110 }
14111
14112 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16_subtile_m) {
14113 TEST_REQUIRES_ARM_NEON;
14114 for (uint32_t m = 1; m <= 1; m++) {
14115 GemmMicrokernelTester()
14116 .mr(1)
14117 .nr(16)
14118 .kr(2)
14119 .sr(1)
14120 .m(m)
14121 .n(16)
14122 .k(16)
14123 .iterations(1)
14124 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14125 }
14126 }
14127
14128 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_eq_16_subtile_n) {
14129 TEST_REQUIRES_ARM_NEON;
14130 for (uint32_t n = 1; n <= 16; n++) {
14131 GemmMicrokernelTester()
14132 .mr(1)
14133 .nr(16)
14134 .kr(2)
14135 .sr(1)
14136 .m(1)
14137 .n(n)
14138 .k(16)
14139 .iterations(1)
14140 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14141 }
14142 }
14143
14144 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_lt_16) {
14145 TEST_REQUIRES_ARM_NEON;
14146 for (size_t k = 1; k < 16; k++) {
14147 GemmMicrokernelTester()
14148 .mr(1)
14149 .nr(16)
14150 .kr(2)
14151 .sr(1)
14152 .m(1)
14153 .n(16)
14154 .k(k)
14155 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14156 }
14157 }
14158
14159 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_lt_16_subtile) {
14160 TEST_REQUIRES_ARM_NEON;
14161 for (size_t k = 1; k < 16; k++) {
14162 for (uint32_t n = 1; n <= 16; n++) {
14163 for (uint32_t m = 1; m <= 1; m++) {
14164 GemmMicrokernelTester()
14165 .mr(1)
14166 .nr(16)
14167 .kr(2)
14168 .sr(1)
14169 .m(m)
14170 .n(n)
14171 .k(k)
14172 .iterations(1)
14173 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14174 }
14175 }
14176 }
14177 }
14178
14179 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_gt_16) {
14180 TEST_REQUIRES_ARM_NEON;
14181 for (size_t k = 17; k < 32; k++) {
14182 GemmMicrokernelTester()
14183 .mr(1)
14184 .nr(16)
14185 .kr(2)
14186 .sr(1)
14187 .m(1)
14188 .n(16)
14189 .k(k)
14190 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14191 }
14192 }
14193
14194 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_gt_16_subtile) {
14195 TEST_REQUIRES_ARM_NEON;
14196 for (size_t k = 17; k < 32; k++) {
14197 for (uint32_t n = 1; n <= 16; n++) {
14198 for (uint32_t m = 1; m <= 1; m++) {
14199 GemmMicrokernelTester()
14200 .mr(1)
14201 .nr(16)
14202 .kr(2)
14203 .sr(1)
14204 .m(m)
14205 .n(n)
14206 .k(k)
14207 .iterations(1)
14208 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14209 }
14210 }
14211 }
14212 }
14213
14214 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_div_16) {
14215 TEST_REQUIRES_ARM_NEON;
14216 for (size_t k = 32; k <= 160; k += 16) {
14217 GemmMicrokernelTester()
14218 .mr(1)
14219 .nr(16)
14220 .kr(2)
14221 .sr(1)
14222 .m(1)
14223 .n(16)
14224 .k(k)
14225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14226 }
14227 }
14228
14229 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, k_div_16_subtile) {
14230 TEST_REQUIRES_ARM_NEON;
14231 for (size_t k = 32; k <= 160; k += 16) {
14232 for (uint32_t n = 1; n <= 16; n++) {
14233 for (uint32_t m = 1; m <= 1; m++) {
14234 GemmMicrokernelTester()
14235 .mr(1)
14236 .nr(16)
14237 .kr(2)
14238 .sr(1)
14239 .m(m)
14240 .n(n)
14241 .k(k)
14242 .iterations(1)
14243 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14244 }
14245 }
14246 }
14247 }
14248
14249 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16) {
14250 TEST_REQUIRES_ARM_NEON;
14251 for (uint32_t n = 17; n < 32; n++) {
14252 for (size_t k = 1; k <= 80; k += 17) {
14253 GemmMicrokernelTester()
14254 .mr(1)
14255 .nr(16)
14256 .kr(2)
14257 .sr(1)
14258 .m(1)
14259 .n(n)
14260 .k(k)
14261 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14262 }
14263 }
14264 }
14265
14266 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16_strided_cn) {
14267 TEST_REQUIRES_ARM_NEON;
14268 for (uint32_t n = 17; n < 32; n++) {
14269 for (size_t k = 1; k <= 80; k += 17) {
14270 GemmMicrokernelTester()
14271 .mr(1)
14272 .nr(16)
14273 .kr(2)
14274 .sr(1)
14275 .m(1)
14276 .n(n)
14277 .k(k)
14278 .cn_stride(19)
14279 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14280 }
14281 }
14282 }
14283
14284 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16_subtile) {
14285 TEST_REQUIRES_ARM_NEON;
14286 for (uint32_t n = 17; n < 32; n++) {
14287 for (size_t k = 1; k <= 80; k += 17) {
14288 for (uint32_t m = 1; m <= 1; m++) {
14289 GemmMicrokernelTester()
14290 .mr(1)
14291 .nr(16)
14292 .kr(2)
14293 .sr(1)
14294 .m(m)
14295 .n(n)
14296 .k(k)
14297 .iterations(1)
14298 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14299 }
14300 }
14301 }
14302 }
14303
14304 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16) {
14305 TEST_REQUIRES_ARM_NEON;
14306 for (uint32_t n = 32; n <= 48; n += 16) {
14307 for (size_t k = 1; k <= 80; k += 17) {
14308 GemmMicrokernelTester()
14309 .mr(1)
14310 .nr(16)
14311 .kr(2)
14312 .sr(1)
14313 .m(1)
14314 .n(n)
14315 .k(k)
14316 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14317 }
14318 }
14319 }
14320
14321 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16_strided_cn) {
14322 TEST_REQUIRES_ARM_NEON;
14323 for (uint32_t n = 32; n <= 48; n += 16) {
14324 for (size_t k = 1; k <= 80; k += 17) {
14325 GemmMicrokernelTester()
14326 .mr(1)
14327 .nr(16)
14328 .kr(2)
14329 .sr(1)
14330 .m(1)
14331 .n(n)
14332 .k(k)
14333 .cn_stride(19)
14334 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14335 }
14336 }
14337 }
14338
14339 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16_subtile) {
14340 TEST_REQUIRES_ARM_NEON;
14341 for (uint32_t n = 32; n <= 48; n += 16) {
14342 for (size_t k = 1; k <= 80; k += 17) {
14343 for (uint32_t m = 1; m <= 1; m++) {
14344 GemmMicrokernelTester()
14345 .mr(1)
14346 .nr(16)
14347 .kr(2)
14348 .sr(1)
14349 .m(m)
14350 .n(n)
14351 .k(k)
14352 .iterations(1)
14353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14354 }
14355 }
14356 }
14357 }
14358
14359 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, small_kernel) {
14360 TEST_REQUIRES_ARM_NEON;
14361 for (size_t k = 1; k <= 80; k += 17) {
14362 GemmMicrokernelTester()
14363 .mr(1)
14364 .nr(16)
14365 .kr(2)
14366 .sr(1)
14367 .m(1)
14368 .n(16)
14369 .k(k)
14370 .ks(3)
14371 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14372 }
14373 }
14374
14375 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, small_kernel_subtile) {
14376 TEST_REQUIRES_ARM_NEON;
14377 for (size_t k = 1; k <= 80; k += 17) {
14378 for (uint32_t n = 1; n <= 16; n++) {
14379 for (uint32_t m = 1; m <= 1; m++) {
14380 GemmMicrokernelTester()
14381 .mr(1)
14382 .nr(16)
14383 .kr(2)
14384 .sr(1)
14385 .m(m)
14386 .n(n)
14387 .k(k)
14388 .ks(3)
14389 .iterations(1)
14390 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14391 }
14392 }
14393 }
14394 }
14395
14396 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_gt_16_small_kernel) {
14397 TEST_REQUIRES_ARM_NEON;
14398 for (uint32_t n = 17; n < 32; n++) {
14399 for (size_t k = 1; k <= 80; k += 17) {
14400 GemmMicrokernelTester()
14401 .mr(1)
14402 .nr(16)
14403 .kr(2)
14404 .sr(1)
14405 .m(1)
14406 .n(n)
14407 .k(k)
14408 .ks(3)
14409 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14410 }
14411 }
14412 }
14413
14414 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, n_div_16_small_kernel) {
14415 TEST_REQUIRES_ARM_NEON;
14416 for (uint32_t n = 32; n <= 48; n += 16) {
14417 for (size_t k = 1; k <= 80; k += 17) {
14418 GemmMicrokernelTester()
14419 .mr(1)
14420 .nr(16)
14421 .kr(2)
14422 .sr(1)
14423 .m(1)
14424 .n(n)
14425 .k(k)
14426 .ks(3)
14427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14428 }
14429 }
14430 }
14431
14432 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, strided_cm_subtile) {
14433 TEST_REQUIRES_ARM_NEON;
14434 for (size_t k = 1; k <= 80; k += 17) {
14435 for (uint32_t n = 1; n <= 16; n++) {
14436 for (uint32_t m = 1; m <= 1; m++) {
14437 GemmMicrokernelTester()
14438 .mr(1)
14439 .nr(16)
14440 .kr(2)
14441 .sr(1)
14442 .m(m)
14443 .n(n)
14444 .k(k)
14445 .cm_stride(19)
14446 .iterations(1)
14447 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14448 }
14449 }
14450 }
14451 }
14452
14453 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, a_offset) {
14454 TEST_REQUIRES_ARM_NEON;
14455 for (size_t k = 1; k <= 80; k += 17) {
14456 GemmMicrokernelTester()
14457 .mr(1)
14458 .nr(16)
14459 .kr(2)
14460 .sr(1)
14461 .m(1)
14462 .n(16)
14463 .k(k)
14464 .ks(3)
14465 .a_offset(83)
14466 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14467 }
14468 }
14469
14470 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, zero) {
14471 TEST_REQUIRES_ARM_NEON;
14472 for (size_t k = 1; k <= 80; k += 17) {
14473 for (uint32_t mz = 0; mz < 1; mz++) {
14474 GemmMicrokernelTester()
14475 .mr(1)
14476 .nr(16)
14477 .kr(2)
14478 .sr(1)
14479 .m(1)
14480 .n(16)
14481 .k(k)
14482 .ks(3)
14483 .a_offset(83)
14484 .zero_index(mz)
14485 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14486 }
14487 }
14488 }
14489
14490 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, qmin) {
14491 TEST_REQUIRES_ARM_NEON;
14492 GemmMicrokernelTester()
14493 .mr(1)
14494 .nr(16)
14495 .kr(2)
14496 .sr(1)
14497 .m(1)
14498 .n(16)
14499 .k(16)
14500 .qmin(128)
14501 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14502 }
14503
14504 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, qmax) {
14505 TEST_REQUIRES_ARM_NEON;
14506 GemmMicrokernelTester()
14507 .mr(1)
14508 .nr(16)
14509 .kr(2)
14510 .sr(1)
14511 .m(1)
14512 .n(16)
14513 .k(16)
14514 .qmax(128)
14515 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14516 }
14517
14518 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C2__NEON_MLAL_DUP, strided_cm) {
14519 TEST_REQUIRES_ARM_NEON;
14520 GemmMicrokernelTester()
14521 .mr(1)
14522 .nr(16)
14523 .kr(2)
14524 .sr(1)
14525 .m(1)
14526 .n(16)
14527 .k(16)
14528 .cm_stride(19)
14529 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c2__neon_mlal_dup, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14530 }
14531#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
14532
14533
14534#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
14535 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_eq_16) {
14536 TEST_REQUIRES_ARM_NEON;
14537 GemmMicrokernelTester()
14538 .mr(2)
14539 .nr(8)
14540 .kr(8)
14541 .sr(1)
14542 .m(2)
14543 .n(8)
14544 .k(16)
14545 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14546 }
14547
14548 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, strided_cn) {
14549 TEST_REQUIRES_ARM_NEON;
14550 GemmMicrokernelTester()
14551 .mr(2)
14552 .nr(8)
14553 .kr(8)
14554 .sr(1)
14555 .m(2)
14556 .n(8)
14557 .k(16)
14558 .cn_stride(11)
14559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14560 }
14561
14562 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_eq_16_subtile) {
14563 TEST_REQUIRES_ARM_NEON;
14564 for (uint32_t n = 1; n <= 8; n++) {
14565 for (uint32_t m = 1; m <= 2; m++) {
14566 GemmMicrokernelTester()
14567 .mr(2)
14568 .nr(8)
14569 .kr(8)
14570 .sr(1)
14571 .m(m)
14572 .n(n)
14573 .k(16)
14574 .iterations(1)
14575 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14576 }
14577 }
14578 }
14579
14580 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_eq_16_subtile_m) {
14581 TEST_REQUIRES_ARM_NEON;
14582 for (uint32_t m = 1; m <= 2; m++) {
14583 GemmMicrokernelTester()
14584 .mr(2)
14585 .nr(8)
14586 .kr(8)
14587 .sr(1)
14588 .m(m)
14589 .n(8)
14590 .k(16)
14591 .iterations(1)
14592 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14593 }
14594 }
14595
14596 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_eq_16_subtile_n) {
14597 TEST_REQUIRES_ARM_NEON;
14598 for (uint32_t n = 1; n <= 8; n++) {
14599 GemmMicrokernelTester()
14600 .mr(2)
14601 .nr(8)
14602 .kr(8)
14603 .sr(1)
14604 .m(2)
14605 .n(n)
14606 .k(16)
14607 .iterations(1)
14608 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14609 }
14610 }
14611
14612 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_lt_16) {
14613 TEST_REQUIRES_ARM_NEON;
14614 for (size_t k = 1; k < 16; k++) {
14615 GemmMicrokernelTester()
14616 .mr(2)
14617 .nr(8)
14618 .kr(8)
14619 .sr(1)
14620 .m(2)
14621 .n(8)
14622 .k(k)
14623 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14624 }
14625 }
14626
14627 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_lt_16_subtile) {
14628 TEST_REQUIRES_ARM_NEON;
14629 for (size_t k = 1; k < 16; k++) {
14630 for (uint32_t n = 1; n <= 8; n++) {
14631 for (uint32_t m = 1; m <= 2; m++) {
14632 GemmMicrokernelTester()
14633 .mr(2)
14634 .nr(8)
14635 .kr(8)
14636 .sr(1)
14637 .m(m)
14638 .n(n)
14639 .k(k)
14640 .iterations(1)
14641 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14642 }
14643 }
14644 }
14645 }
14646
14647 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_gt_16) {
14648 TEST_REQUIRES_ARM_NEON;
14649 for (size_t k = 17; k < 32; k++) {
14650 GemmMicrokernelTester()
14651 .mr(2)
14652 .nr(8)
14653 .kr(8)
14654 .sr(1)
14655 .m(2)
14656 .n(8)
14657 .k(k)
14658 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14659 }
14660 }
14661
14662 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_gt_16_subtile) {
14663 TEST_REQUIRES_ARM_NEON;
14664 for (size_t k = 17; k < 32; k++) {
14665 for (uint32_t n = 1; n <= 8; n++) {
14666 for (uint32_t m = 1; m <= 2; m++) {
14667 GemmMicrokernelTester()
14668 .mr(2)
14669 .nr(8)
14670 .kr(8)
14671 .sr(1)
14672 .m(m)
14673 .n(n)
14674 .k(k)
14675 .iterations(1)
14676 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14677 }
14678 }
14679 }
14680 }
14681
14682 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_div_16) {
14683 TEST_REQUIRES_ARM_NEON;
14684 for (size_t k = 32; k <= 160; k += 16) {
14685 GemmMicrokernelTester()
14686 .mr(2)
14687 .nr(8)
14688 .kr(8)
14689 .sr(1)
14690 .m(2)
14691 .n(8)
14692 .k(k)
14693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14694 }
14695 }
14696
14697 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, k_div_16_subtile) {
14698 TEST_REQUIRES_ARM_NEON;
14699 for (size_t k = 32; k <= 160; k += 16) {
14700 for (uint32_t n = 1; n <= 8; n++) {
14701 for (uint32_t m = 1; m <= 2; m++) {
14702 GemmMicrokernelTester()
14703 .mr(2)
14704 .nr(8)
14705 .kr(8)
14706 .sr(1)
14707 .m(m)
14708 .n(n)
14709 .k(k)
14710 .iterations(1)
14711 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14712 }
14713 }
14714 }
14715 }
14716
14717 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_gt_8) {
14718 TEST_REQUIRES_ARM_NEON;
14719 for (uint32_t n = 9; n < 16; n++) {
14720 for (size_t k = 1; k <= 80; k += 17) {
14721 GemmMicrokernelTester()
14722 .mr(2)
14723 .nr(8)
14724 .kr(8)
14725 .sr(1)
14726 .m(2)
14727 .n(n)
14728 .k(k)
14729 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14730 }
14731 }
14732 }
14733
14734 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_gt_8_strided_cn) {
14735 TEST_REQUIRES_ARM_NEON;
14736 for (uint32_t n = 9; n < 16; n++) {
14737 for (size_t k = 1; k <= 80; k += 17) {
14738 GemmMicrokernelTester()
14739 .mr(2)
14740 .nr(8)
14741 .kr(8)
14742 .sr(1)
14743 .m(2)
14744 .n(n)
14745 .k(k)
14746 .cn_stride(11)
14747 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14748 }
14749 }
14750 }
14751
14752 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_gt_8_subtile) {
14753 TEST_REQUIRES_ARM_NEON;
14754 for (uint32_t n = 9; n < 16; n++) {
14755 for (size_t k = 1; k <= 80; k += 17) {
14756 for (uint32_t m = 1; m <= 2; m++) {
14757 GemmMicrokernelTester()
14758 .mr(2)
14759 .nr(8)
14760 .kr(8)
14761 .sr(1)
14762 .m(m)
14763 .n(n)
14764 .k(k)
14765 .iterations(1)
14766 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14767 }
14768 }
14769 }
14770 }
14771
14772 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_div_8) {
14773 TEST_REQUIRES_ARM_NEON;
14774 for (uint32_t n = 16; n <= 24; n += 8) {
14775 for (size_t k = 1; k <= 80; k += 17) {
14776 GemmMicrokernelTester()
14777 .mr(2)
14778 .nr(8)
14779 .kr(8)
14780 .sr(1)
14781 .m(2)
14782 .n(n)
14783 .k(k)
14784 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14785 }
14786 }
14787 }
14788
14789 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_div_8_strided_cn) {
14790 TEST_REQUIRES_ARM_NEON;
14791 for (uint32_t n = 16; n <= 24; n += 8) {
14792 for (size_t k = 1; k <= 80; k += 17) {
14793 GemmMicrokernelTester()
14794 .mr(2)
14795 .nr(8)
14796 .kr(8)
14797 .sr(1)
14798 .m(2)
14799 .n(n)
14800 .k(k)
14801 .cn_stride(11)
14802 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14803 }
14804 }
14805 }
14806
14807 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_div_8_subtile) {
14808 TEST_REQUIRES_ARM_NEON;
14809 for (uint32_t n = 16; n <= 24; n += 8) {
14810 for (size_t k = 1; k <= 80; k += 17) {
14811 for (uint32_t m = 1; m <= 2; m++) {
14812 GemmMicrokernelTester()
14813 .mr(2)
14814 .nr(8)
14815 .kr(8)
14816 .sr(1)
14817 .m(m)
14818 .n(n)
14819 .k(k)
14820 .iterations(1)
14821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14822 }
14823 }
14824 }
14825 }
14826
14827 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, small_kernel) {
14828 TEST_REQUIRES_ARM_NEON;
14829 for (size_t k = 1; k <= 80; k += 17) {
14830 GemmMicrokernelTester()
14831 .mr(2)
14832 .nr(8)
14833 .kr(8)
14834 .sr(1)
14835 .m(2)
14836 .n(8)
14837 .k(k)
14838 .ks(3)
14839 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14840 }
14841 }
14842
14843 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, small_kernel_subtile) {
14844 TEST_REQUIRES_ARM_NEON;
14845 for (size_t k = 1; k <= 80; k += 17) {
14846 for (uint32_t n = 1; n <= 8; n++) {
14847 for (uint32_t m = 1; m <= 2; m++) {
14848 GemmMicrokernelTester()
14849 .mr(2)
14850 .nr(8)
14851 .kr(8)
14852 .sr(1)
14853 .m(m)
14854 .n(n)
14855 .k(k)
14856 .ks(3)
14857 .iterations(1)
14858 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14859 }
14860 }
14861 }
14862 }
14863
14864 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_gt_8_small_kernel) {
14865 TEST_REQUIRES_ARM_NEON;
14866 for (uint32_t n = 9; n < 16; n++) {
14867 for (size_t k = 1; k <= 80; k += 17) {
14868 GemmMicrokernelTester()
14869 .mr(2)
14870 .nr(8)
14871 .kr(8)
14872 .sr(1)
14873 .m(2)
14874 .n(n)
14875 .k(k)
14876 .ks(3)
14877 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14878 }
14879 }
14880 }
14881
14882 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, n_div_8_small_kernel) {
14883 TEST_REQUIRES_ARM_NEON;
14884 for (uint32_t n = 16; n <= 24; n += 8) {
14885 for (size_t k = 1; k <= 80; k += 17) {
14886 GemmMicrokernelTester()
14887 .mr(2)
14888 .nr(8)
14889 .kr(8)
14890 .sr(1)
14891 .m(2)
14892 .n(n)
14893 .k(k)
14894 .ks(3)
14895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14896 }
14897 }
14898 }
14899
14900 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, strided_cm_subtile) {
14901 TEST_REQUIRES_ARM_NEON;
14902 for (size_t k = 1; k <= 80; k += 17) {
14903 for (uint32_t n = 1; n <= 8; n++) {
14904 for (uint32_t m = 1; m <= 2; m++) {
14905 GemmMicrokernelTester()
14906 .mr(2)
14907 .nr(8)
14908 .kr(8)
14909 .sr(1)
14910 .m(m)
14911 .n(n)
14912 .k(k)
14913 .cm_stride(11)
14914 .iterations(1)
14915 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14916 }
14917 }
14918 }
14919 }
14920
14921 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, a_offset) {
14922 TEST_REQUIRES_ARM_NEON;
14923 for (size_t k = 1; k <= 80; k += 17) {
14924 GemmMicrokernelTester()
14925 .mr(2)
14926 .nr(8)
14927 .kr(8)
14928 .sr(1)
14929 .m(2)
14930 .n(8)
14931 .k(k)
14932 .ks(3)
14933 .a_offset(163)
14934 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14935 }
14936 }
14937
14938 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, zero) {
14939 TEST_REQUIRES_ARM_NEON;
14940 for (size_t k = 1; k <= 80; k += 17) {
14941 for (uint32_t mz = 0; mz < 2; mz++) {
14942 GemmMicrokernelTester()
14943 .mr(2)
14944 .nr(8)
14945 .kr(8)
14946 .sr(1)
14947 .m(2)
14948 .n(8)
14949 .k(k)
14950 .ks(3)
14951 .a_offset(163)
14952 .zero_index(mz)
14953 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14954 }
14955 }
14956 }
14957
14958 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, qmin) {
14959 TEST_REQUIRES_ARM_NEON;
14960 GemmMicrokernelTester()
14961 .mr(2)
14962 .nr(8)
14963 .kr(8)
14964 .sr(1)
14965 .m(2)
14966 .n(8)
14967 .k(16)
14968 .qmin(128)
14969 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14970 }
14971
14972 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, qmax) {
14973 TEST_REQUIRES_ARM_NEON;
14974 GemmMicrokernelTester()
14975 .mr(2)
14976 .nr(8)
14977 .kr(8)
14978 .sr(1)
14979 .m(2)
14980 .n(8)
14981 .k(16)
14982 .qmax(128)
14983 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14984 }
14985
14986 TEST(QS8_IGEMM_MINMAX_RNDNU_2X8C8__AARCH64_NEON_MLAL, strided_cm) {
14987 TEST_REQUIRES_ARM_NEON;
14988 GemmMicrokernelTester()
14989 .mr(2)
14990 .nr(8)
14991 .kr(8)
14992 .sr(1)
14993 .m(2)
14994 .n(8)
14995 .k(16)
14996 .cm_stride(11)
14997 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
14998 }
14999#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
15000
15001
15002#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
15003 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_eq_16) {
15004 TEST_REQUIRES_ARM_NEON;
15005 GemmMicrokernelTester()
15006 .mr(1)
15007 .nr(8)
15008 .kr(8)
15009 .sr(1)
15010 .m(1)
15011 .n(8)
15012 .k(16)
15013 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15014 }
15015
15016 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, strided_cn) {
15017 TEST_REQUIRES_ARM_NEON;
15018 GemmMicrokernelTester()
15019 .mr(1)
15020 .nr(8)
15021 .kr(8)
15022 .sr(1)
15023 .m(1)
15024 .n(8)
15025 .k(16)
15026 .cn_stride(11)
15027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15028 }
15029
15030 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_eq_16_subtile) {
15031 TEST_REQUIRES_ARM_NEON;
15032 for (uint32_t n = 1; n <= 8; n++) {
15033 for (uint32_t m = 1; m <= 1; m++) {
15034 GemmMicrokernelTester()
15035 .mr(1)
15036 .nr(8)
15037 .kr(8)
15038 .sr(1)
15039 .m(m)
15040 .n(n)
15041 .k(16)
15042 .iterations(1)
15043 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15044 }
15045 }
15046 }
15047
15048 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_eq_16_subtile_m) {
15049 TEST_REQUIRES_ARM_NEON;
15050 for (uint32_t m = 1; m <= 1; m++) {
15051 GemmMicrokernelTester()
15052 .mr(1)
15053 .nr(8)
15054 .kr(8)
15055 .sr(1)
15056 .m(m)
15057 .n(8)
15058 .k(16)
15059 .iterations(1)
15060 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15061 }
15062 }
15063
15064 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_eq_16_subtile_n) {
15065 TEST_REQUIRES_ARM_NEON;
15066 for (uint32_t n = 1; n <= 8; n++) {
15067 GemmMicrokernelTester()
15068 .mr(1)
15069 .nr(8)
15070 .kr(8)
15071 .sr(1)
15072 .m(1)
15073 .n(n)
15074 .k(16)
15075 .iterations(1)
15076 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15077 }
15078 }
15079
15080 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_lt_16) {
15081 TEST_REQUIRES_ARM_NEON;
15082 for (size_t k = 1; k < 16; k++) {
15083 GemmMicrokernelTester()
15084 .mr(1)
15085 .nr(8)
15086 .kr(8)
15087 .sr(1)
15088 .m(1)
15089 .n(8)
15090 .k(k)
15091 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15092 }
15093 }
15094
15095 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_lt_16_subtile) {
15096 TEST_REQUIRES_ARM_NEON;
15097 for (size_t k = 1; k < 16; k++) {
15098 for (uint32_t n = 1; n <= 8; n++) {
15099 for (uint32_t m = 1; m <= 1; m++) {
15100 GemmMicrokernelTester()
15101 .mr(1)
15102 .nr(8)
15103 .kr(8)
15104 .sr(1)
15105 .m(m)
15106 .n(n)
15107 .k(k)
15108 .iterations(1)
15109 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15110 }
15111 }
15112 }
15113 }
15114
15115 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_gt_16) {
15116 TEST_REQUIRES_ARM_NEON;
15117 for (size_t k = 17; k < 32; k++) {
15118 GemmMicrokernelTester()
15119 .mr(1)
15120 .nr(8)
15121 .kr(8)
15122 .sr(1)
15123 .m(1)
15124 .n(8)
15125 .k(k)
15126 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15127 }
15128 }
15129
15130 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_gt_16_subtile) {
15131 TEST_REQUIRES_ARM_NEON;
15132 for (size_t k = 17; k < 32; k++) {
15133 for (uint32_t n = 1; n <= 8; n++) {
15134 for (uint32_t m = 1; m <= 1; m++) {
15135 GemmMicrokernelTester()
15136 .mr(1)
15137 .nr(8)
15138 .kr(8)
15139 .sr(1)
15140 .m(m)
15141 .n(n)
15142 .k(k)
15143 .iterations(1)
15144 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15145 }
15146 }
15147 }
15148 }
15149
15150 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_div_16) {
15151 TEST_REQUIRES_ARM_NEON;
15152 for (size_t k = 32; k <= 160; k += 16) {
15153 GemmMicrokernelTester()
15154 .mr(1)
15155 .nr(8)
15156 .kr(8)
15157 .sr(1)
15158 .m(1)
15159 .n(8)
15160 .k(k)
15161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15162 }
15163 }
15164
15165 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, k_div_16_subtile) {
15166 TEST_REQUIRES_ARM_NEON;
15167 for (size_t k = 32; k <= 160; k += 16) {
15168 for (uint32_t n = 1; n <= 8; n++) {
15169 for (uint32_t m = 1; m <= 1; m++) {
15170 GemmMicrokernelTester()
15171 .mr(1)
15172 .nr(8)
15173 .kr(8)
15174 .sr(1)
15175 .m(m)
15176 .n(n)
15177 .k(k)
15178 .iterations(1)
15179 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15180 }
15181 }
15182 }
15183 }
15184
15185 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_gt_8) {
15186 TEST_REQUIRES_ARM_NEON;
15187 for (uint32_t n = 9; n < 16; n++) {
15188 for (size_t k = 1; k <= 80; k += 17) {
15189 GemmMicrokernelTester()
15190 .mr(1)
15191 .nr(8)
15192 .kr(8)
15193 .sr(1)
15194 .m(1)
15195 .n(n)
15196 .k(k)
15197 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15198 }
15199 }
15200 }
15201
15202 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_gt_8_strided_cn) {
15203 TEST_REQUIRES_ARM_NEON;
15204 for (uint32_t n = 9; n < 16; n++) {
15205 for (size_t k = 1; k <= 80; k += 17) {
15206 GemmMicrokernelTester()
15207 .mr(1)
15208 .nr(8)
15209 .kr(8)
15210 .sr(1)
15211 .m(1)
15212 .n(n)
15213 .k(k)
15214 .cn_stride(11)
15215 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15216 }
15217 }
15218 }
15219
15220 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_gt_8_subtile) {
15221 TEST_REQUIRES_ARM_NEON;
15222 for (uint32_t n = 9; n < 16; n++) {
15223 for (size_t k = 1; k <= 80; k += 17) {
15224 for (uint32_t m = 1; m <= 1; m++) {
15225 GemmMicrokernelTester()
15226 .mr(1)
15227 .nr(8)
15228 .kr(8)
15229 .sr(1)
15230 .m(m)
15231 .n(n)
15232 .k(k)
15233 .iterations(1)
15234 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15235 }
15236 }
15237 }
15238 }
15239
15240 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_div_8) {
15241 TEST_REQUIRES_ARM_NEON;
15242 for (uint32_t n = 16; n <= 24; n += 8) {
15243 for (size_t k = 1; k <= 80; k += 17) {
15244 GemmMicrokernelTester()
15245 .mr(1)
15246 .nr(8)
15247 .kr(8)
15248 .sr(1)
15249 .m(1)
15250 .n(n)
15251 .k(k)
15252 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15253 }
15254 }
15255 }
15256
15257 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_div_8_strided_cn) {
15258 TEST_REQUIRES_ARM_NEON;
15259 for (uint32_t n = 16; n <= 24; n += 8) {
15260 for (size_t k = 1; k <= 80; k += 17) {
15261 GemmMicrokernelTester()
15262 .mr(1)
15263 .nr(8)
15264 .kr(8)
15265 .sr(1)
15266 .m(1)
15267 .n(n)
15268 .k(k)
15269 .cn_stride(11)
15270 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15271 }
15272 }
15273 }
15274
15275 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_div_8_subtile) {
15276 TEST_REQUIRES_ARM_NEON;
15277 for (uint32_t n = 16; n <= 24; n += 8) {
15278 for (size_t k = 1; k <= 80; k += 17) {
15279 for (uint32_t m = 1; m <= 1; m++) {
15280 GemmMicrokernelTester()
15281 .mr(1)
15282 .nr(8)
15283 .kr(8)
15284 .sr(1)
15285 .m(m)
15286 .n(n)
15287 .k(k)
15288 .iterations(1)
15289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15290 }
15291 }
15292 }
15293 }
15294
15295 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, small_kernel) {
15296 TEST_REQUIRES_ARM_NEON;
15297 for (size_t k = 1; k <= 80; k += 17) {
15298 GemmMicrokernelTester()
15299 .mr(1)
15300 .nr(8)
15301 .kr(8)
15302 .sr(1)
15303 .m(1)
15304 .n(8)
15305 .k(k)
15306 .ks(3)
15307 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15308 }
15309 }
15310
15311 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, small_kernel_subtile) {
15312 TEST_REQUIRES_ARM_NEON;
15313 for (size_t k = 1; k <= 80; k += 17) {
15314 for (uint32_t n = 1; n <= 8; n++) {
15315 for (uint32_t m = 1; m <= 1; m++) {
15316 GemmMicrokernelTester()
15317 .mr(1)
15318 .nr(8)
15319 .kr(8)
15320 .sr(1)
15321 .m(m)
15322 .n(n)
15323 .k(k)
15324 .ks(3)
15325 .iterations(1)
15326 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15327 }
15328 }
15329 }
15330 }
15331
15332 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_gt_8_small_kernel) {
15333 TEST_REQUIRES_ARM_NEON;
15334 for (uint32_t n = 9; n < 16; n++) {
15335 for (size_t k = 1; k <= 80; k += 17) {
15336 GemmMicrokernelTester()
15337 .mr(1)
15338 .nr(8)
15339 .kr(8)
15340 .sr(1)
15341 .m(1)
15342 .n(n)
15343 .k(k)
15344 .ks(3)
15345 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15346 }
15347 }
15348 }
15349
15350 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, n_div_8_small_kernel) {
15351 TEST_REQUIRES_ARM_NEON;
15352 for (uint32_t n = 16; n <= 24; n += 8) {
15353 for (size_t k = 1; k <= 80; k += 17) {
15354 GemmMicrokernelTester()
15355 .mr(1)
15356 .nr(8)
15357 .kr(8)
15358 .sr(1)
15359 .m(1)
15360 .n(n)
15361 .k(k)
15362 .ks(3)
15363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15364 }
15365 }
15366 }
15367
15368 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, strided_cm_subtile) {
15369 TEST_REQUIRES_ARM_NEON;
15370 for (size_t k = 1; k <= 80; k += 17) {
15371 for (uint32_t n = 1; n <= 8; n++) {
15372 for (uint32_t m = 1; m <= 1; m++) {
15373 GemmMicrokernelTester()
15374 .mr(1)
15375 .nr(8)
15376 .kr(8)
15377 .sr(1)
15378 .m(m)
15379 .n(n)
15380 .k(k)
15381 .cm_stride(11)
15382 .iterations(1)
15383 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15384 }
15385 }
15386 }
15387 }
15388
15389 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, a_offset) {
15390 TEST_REQUIRES_ARM_NEON;
15391 for (size_t k = 1; k <= 80; k += 17) {
15392 GemmMicrokernelTester()
15393 .mr(1)
15394 .nr(8)
15395 .kr(8)
15396 .sr(1)
15397 .m(1)
15398 .n(8)
15399 .k(k)
15400 .ks(3)
15401 .a_offset(83)
15402 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15403 }
15404 }
15405
15406 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, zero) {
15407 TEST_REQUIRES_ARM_NEON;
15408 for (size_t k = 1; k <= 80; k += 17) {
15409 for (uint32_t mz = 0; mz < 1; mz++) {
15410 GemmMicrokernelTester()
15411 .mr(1)
15412 .nr(8)
15413 .kr(8)
15414 .sr(1)
15415 .m(1)
15416 .n(8)
15417 .k(k)
15418 .ks(3)
15419 .a_offset(83)
15420 .zero_index(mz)
15421 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15422 }
15423 }
15424 }
15425
15426 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, qmin) {
15427 TEST_REQUIRES_ARM_NEON;
15428 GemmMicrokernelTester()
15429 .mr(1)
15430 .nr(8)
15431 .kr(8)
15432 .sr(1)
15433 .m(1)
15434 .n(8)
15435 .k(16)
15436 .qmin(128)
15437 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15438 }
15439
15440 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, qmax) {
15441 TEST_REQUIRES_ARM_NEON;
15442 GemmMicrokernelTester()
15443 .mr(1)
15444 .nr(8)
15445 .kr(8)
15446 .sr(1)
15447 .m(1)
15448 .n(8)
15449 .k(16)
15450 .qmax(128)
15451 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15452 }
15453
15454 TEST(QS8_IGEMM_MINMAX_RNDNU_1X8C8__AARCH64_NEON_MLAL, strided_cm) {
15455 TEST_REQUIRES_ARM_NEON;
15456 GemmMicrokernelTester()
15457 .mr(1)
15458 .nr(8)
15459 .kr(8)
15460 .sr(1)
15461 .m(1)
15462 .n(8)
15463 .k(16)
15464 .cm_stride(11)
15465 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x8c8__aarch64_neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15466 }
15467#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
15468
15469
15470#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
15471 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_eq_8) {
15472 TEST_REQUIRES_ARM_NEON;
15473 GemmMicrokernelTester()
15474 .mr(4)
15475 .nr(16)
15476 .kr(1)
15477 .sr(1)
15478 .m(4)
15479 .n(16)
15480 .k(8)
15481 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15482 }
15483
15484 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, strided_cn) {
15485 TEST_REQUIRES_ARM_NEON;
15486 GemmMicrokernelTester()
15487 .mr(4)
15488 .nr(16)
15489 .kr(1)
15490 .sr(1)
15491 .m(4)
15492 .n(16)
15493 .k(8)
15494 .cn_stride(19)
15495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15496 }
15497
15498 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_eq_8_subtile) {
15499 TEST_REQUIRES_ARM_NEON;
15500 for (uint32_t n = 1; n <= 16; n++) {
15501 for (uint32_t m = 1; m <= 4; m++) {
15502 GemmMicrokernelTester()
15503 .mr(4)
15504 .nr(16)
15505 .kr(1)
15506 .sr(1)
15507 .m(m)
15508 .n(n)
15509 .k(8)
15510 .iterations(1)
15511 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15512 }
15513 }
15514 }
15515
15516 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_eq_8_subtile_m) {
15517 TEST_REQUIRES_ARM_NEON;
15518 for (uint32_t m = 1; m <= 4; m++) {
15519 GemmMicrokernelTester()
15520 .mr(4)
15521 .nr(16)
15522 .kr(1)
15523 .sr(1)
15524 .m(m)
15525 .n(16)
15526 .k(8)
15527 .iterations(1)
15528 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15529 }
15530 }
15531
15532 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_eq_8_subtile_n) {
15533 TEST_REQUIRES_ARM_NEON;
15534 for (uint32_t n = 1; n <= 16; n++) {
15535 GemmMicrokernelTester()
15536 .mr(4)
15537 .nr(16)
15538 .kr(1)
15539 .sr(1)
15540 .m(4)
15541 .n(n)
15542 .k(8)
15543 .iterations(1)
15544 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15545 }
15546 }
15547
15548 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_lt_8) {
15549 TEST_REQUIRES_ARM_NEON;
15550 for (size_t k = 1; k < 8; k++) {
15551 GemmMicrokernelTester()
15552 .mr(4)
15553 .nr(16)
15554 .kr(1)
15555 .sr(1)
15556 .m(4)
15557 .n(16)
15558 .k(k)
15559 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15560 }
15561 }
15562
15563 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_lt_8_subtile) {
15564 TEST_REQUIRES_ARM_NEON;
15565 for (size_t k = 1; k < 8; k++) {
15566 for (uint32_t n = 1; n <= 16; n++) {
15567 for (uint32_t m = 1; m <= 4; m++) {
15568 GemmMicrokernelTester()
15569 .mr(4)
15570 .nr(16)
15571 .kr(1)
15572 .sr(1)
15573 .m(m)
15574 .n(n)
15575 .k(k)
15576 .iterations(1)
15577 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15578 }
15579 }
15580 }
15581 }
15582
15583 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_gt_8) {
15584 TEST_REQUIRES_ARM_NEON;
15585 for (size_t k = 9; k < 16; k++) {
15586 GemmMicrokernelTester()
15587 .mr(4)
15588 .nr(16)
15589 .kr(1)
15590 .sr(1)
15591 .m(4)
15592 .n(16)
15593 .k(k)
15594 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15595 }
15596 }
15597
15598 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_gt_8_subtile) {
15599 TEST_REQUIRES_ARM_NEON;
15600 for (size_t k = 9; k < 16; k++) {
15601 for (uint32_t n = 1; n <= 16; n++) {
15602 for (uint32_t m = 1; m <= 4; m++) {
15603 GemmMicrokernelTester()
15604 .mr(4)
15605 .nr(16)
15606 .kr(1)
15607 .sr(1)
15608 .m(m)
15609 .n(n)
15610 .k(k)
15611 .iterations(1)
15612 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15613 }
15614 }
15615 }
15616 }
15617
15618 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_div_8) {
15619 TEST_REQUIRES_ARM_NEON;
15620 for (size_t k = 16; k <= 80; k += 8) {
15621 GemmMicrokernelTester()
15622 .mr(4)
15623 .nr(16)
15624 .kr(1)
15625 .sr(1)
15626 .m(4)
15627 .n(16)
15628 .k(k)
15629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15630 }
15631 }
15632
15633 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, k_div_8_subtile) {
15634 TEST_REQUIRES_ARM_NEON;
15635 for (size_t k = 16; k <= 80; k += 8) {
15636 for (uint32_t n = 1; n <= 16; n++) {
15637 for (uint32_t m = 1; m <= 4; m++) {
15638 GemmMicrokernelTester()
15639 .mr(4)
15640 .nr(16)
15641 .kr(1)
15642 .sr(1)
15643 .m(m)
15644 .n(n)
15645 .k(k)
15646 .iterations(1)
15647 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15648 }
15649 }
15650 }
15651 }
15652
15653 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_gt_16) {
15654 TEST_REQUIRES_ARM_NEON;
15655 for (uint32_t n = 17; n < 32; n++) {
15656 for (size_t k = 1; k <= 40; k += 9) {
15657 GemmMicrokernelTester()
15658 .mr(4)
15659 .nr(16)
15660 .kr(1)
15661 .sr(1)
15662 .m(4)
15663 .n(n)
15664 .k(k)
15665 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15666 }
15667 }
15668 }
15669
15670 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_gt_16_strided_cn) {
15671 TEST_REQUIRES_ARM_NEON;
15672 for (uint32_t n = 17; n < 32; n++) {
15673 for (size_t k = 1; k <= 40; k += 9) {
15674 GemmMicrokernelTester()
15675 .mr(4)
15676 .nr(16)
15677 .kr(1)
15678 .sr(1)
15679 .m(4)
15680 .n(n)
15681 .k(k)
15682 .cn_stride(19)
15683 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15684 }
15685 }
15686 }
15687
15688 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_gt_16_subtile) {
15689 TEST_REQUIRES_ARM_NEON;
15690 for (uint32_t n = 17; n < 32; n++) {
15691 for (size_t k = 1; k <= 40; k += 9) {
15692 for (uint32_t m = 1; m <= 4; m++) {
15693 GemmMicrokernelTester()
15694 .mr(4)
15695 .nr(16)
15696 .kr(1)
15697 .sr(1)
15698 .m(m)
15699 .n(n)
15700 .k(k)
15701 .iterations(1)
15702 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15703 }
15704 }
15705 }
15706 }
15707
15708 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_div_16) {
15709 TEST_REQUIRES_ARM_NEON;
15710 for (uint32_t n = 32; n <= 48; n += 16) {
15711 for (size_t k = 1; k <= 40; k += 9) {
15712 GemmMicrokernelTester()
15713 .mr(4)
15714 .nr(16)
15715 .kr(1)
15716 .sr(1)
15717 .m(4)
15718 .n(n)
15719 .k(k)
15720 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15721 }
15722 }
15723 }
15724
15725 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_div_16_strided_cn) {
15726 TEST_REQUIRES_ARM_NEON;
15727 for (uint32_t n = 32; n <= 48; n += 16) {
15728 for (size_t k = 1; k <= 40; k += 9) {
15729 GemmMicrokernelTester()
15730 .mr(4)
15731 .nr(16)
15732 .kr(1)
15733 .sr(1)
15734 .m(4)
15735 .n(n)
15736 .k(k)
15737 .cn_stride(19)
15738 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15739 }
15740 }
15741 }
15742
15743 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_div_16_subtile) {
15744 TEST_REQUIRES_ARM_NEON;
15745 for (uint32_t n = 32; n <= 48; n += 16) {
15746 for (size_t k = 1; k <= 40; k += 9) {
15747 for (uint32_t m = 1; m <= 4; m++) {
15748 GemmMicrokernelTester()
15749 .mr(4)
15750 .nr(16)
15751 .kr(1)
15752 .sr(1)
15753 .m(m)
15754 .n(n)
15755 .k(k)
15756 .iterations(1)
15757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15758 }
15759 }
15760 }
15761 }
15762
15763 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, small_kernel) {
15764 TEST_REQUIRES_ARM_NEON;
15765 for (size_t k = 1; k <= 40; k += 9) {
15766 GemmMicrokernelTester()
15767 .mr(4)
15768 .nr(16)
15769 .kr(1)
15770 .sr(1)
15771 .m(4)
15772 .n(16)
15773 .k(k)
15774 .ks(3)
15775 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15776 }
15777 }
15778
15779 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, small_kernel_subtile) {
15780 TEST_REQUIRES_ARM_NEON;
15781 for (size_t k = 1; k <= 40; k += 9) {
15782 for (uint32_t n = 1; n <= 16; n++) {
15783 for (uint32_t m = 1; m <= 4; m++) {
15784 GemmMicrokernelTester()
15785 .mr(4)
15786 .nr(16)
15787 .kr(1)
15788 .sr(1)
15789 .m(m)
15790 .n(n)
15791 .k(k)
15792 .ks(3)
15793 .iterations(1)
15794 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15795 }
15796 }
15797 }
15798 }
15799
15800 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_gt_16_small_kernel) {
15801 TEST_REQUIRES_ARM_NEON;
15802 for (uint32_t n = 17; n < 32; n++) {
15803 for (size_t k = 1; k <= 40; k += 9) {
15804 GemmMicrokernelTester()
15805 .mr(4)
15806 .nr(16)
15807 .kr(1)
15808 .sr(1)
15809 .m(4)
15810 .n(n)
15811 .k(k)
15812 .ks(3)
15813 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15814 }
15815 }
15816 }
15817
15818 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, n_div_16_small_kernel) {
15819 TEST_REQUIRES_ARM_NEON;
15820 for (uint32_t n = 32; n <= 48; n += 16) {
15821 for (size_t k = 1; k <= 40; k += 9) {
15822 GemmMicrokernelTester()
15823 .mr(4)
15824 .nr(16)
15825 .kr(1)
15826 .sr(1)
15827 .m(4)
15828 .n(n)
15829 .k(k)
15830 .ks(3)
15831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15832 }
15833 }
15834 }
15835
15836 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, strided_cm_subtile) {
15837 TEST_REQUIRES_ARM_NEON;
15838 for (size_t k = 1; k <= 40; k += 9) {
15839 for (uint32_t n = 1; n <= 16; n++) {
15840 for (uint32_t m = 1; m <= 4; m++) {
15841 GemmMicrokernelTester()
15842 .mr(4)
15843 .nr(16)
15844 .kr(1)
15845 .sr(1)
15846 .m(m)
15847 .n(n)
15848 .k(k)
15849 .cm_stride(19)
15850 .iterations(1)
15851 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15852 }
15853 }
15854 }
15855 }
15856
15857 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, a_offset) {
15858 TEST_REQUIRES_ARM_NEON;
15859 for (size_t k = 1; k <= 40; k += 9) {
15860 GemmMicrokernelTester()
15861 .mr(4)
15862 .nr(16)
15863 .kr(1)
15864 .sr(1)
15865 .m(4)
15866 .n(16)
15867 .k(k)
15868 .ks(3)
15869 .a_offset(163)
15870 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15871 }
15872 }
15873
15874 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, zero) {
15875 TEST_REQUIRES_ARM_NEON;
15876 for (size_t k = 1; k <= 40; k += 9) {
15877 for (uint32_t mz = 0; mz < 4; mz++) {
15878 GemmMicrokernelTester()
15879 .mr(4)
15880 .nr(16)
15881 .kr(1)
15882 .sr(1)
15883 .m(4)
15884 .n(16)
15885 .k(k)
15886 .ks(3)
15887 .a_offset(163)
15888 .zero_index(mz)
15889 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15890 }
15891 }
15892 }
15893
15894 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, qmin) {
15895 TEST_REQUIRES_ARM_NEON;
15896 GemmMicrokernelTester()
15897 .mr(4)
15898 .nr(16)
15899 .kr(1)
15900 .sr(1)
15901 .m(4)
15902 .n(16)
15903 .k(8)
15904 .qmin(128)
15905 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15906 }
15907
15908 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, qmax) {
15909 TEST_REQUIRES_ARM_NEON;
15910 GemmMicrokernelTester()
15911 .mr(4)
15912 .nr(16)
15913 .kr(1)
15914 .sr(1)
15915 .m(4)
15916 .n(16)
15917 .k(8)
15918 .qmax(128)
15919 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15920 }
15921
15922 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__AARCH64_NEON_MLAL_LANE_CORTEX_A53, strided_cm) {
15923 TEST_REQUIRES_ARM_NEON;
15924 GemmMicrokernelTester()
15925 .mr(4)
15926 .nr(16)
15927 .kr(1)
15928 .sr(1)
15929 .m(4)
15930 .n(16)
15931 .k(8)
15932 .cm_stride(19)
15933 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15934 }
15935#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
15936
15937
15938#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
15939 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_eq_8) {
15940 TEST_REQUIRES_ARM_NEON_DOT;
15941 GemmMicrokernelTester()
15942 .mr(4)
15943 .nr(16)
15944 .kr(4)
15945 .sr(1)
15946 .m(4)
15947 .n(16)
15948 .k(8)
15949 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15950 }
15951
15952 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, strided_cn) {
15953 TEST_REQUIRES_ARM_NEON_DOT;
15954 GemmMicrokernelTester()
15955 .mr(4)
15956 .nr(16)
15957 .kr(4)
15958 .sr(1)
15959 .m(4)
15960 .n(16)
15961 .k(8)
15962 .cn_stride(19)
15963 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15964 }
15965
15966 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_eq_8_subtile) {
15967 TEST_REQUIRES_ARM_NEON_DOT;
15968 for (uint32_t n = 1; n <= 16; n++) {
15969 for (uint32_t m = 1; m <= 4; m++) {
15970 GemmMicrokernelTester()
15971 .mr(4)
15972 .nr(16)
15973 .kr(4)
15974 .sr(1)
15975 .m(m)
15976 .n(n)
15977 .k(8)
15978 .iterations(1)
15979 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15980 }
15981 }
15982 }
15983
15984 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_eq_8_subtile_m) {
15985 TEST_REQUIRES_ARM_NEON_DOT;
15986 for (uint32_t m = 1; m <= 4; m++) {
15987 GemmMicrokernelTester()
15988 .mr(4)
15989 .nr(16)
15990 .kr(4)
15991 .sr(1)
15992 .m(m)
15993 .n(16)
15994 .k(8)
15995 .iterations(1)
15996 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
15997 }
15998 }
15999
16000 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_eq_8_subtile_n) {
16001 TEST_REQUIRES_ARM_NEON_DOT;
16002 for (uint32_t n = 1; n <= 16; n++) {
16003 GemmMicrokernelTester()
16004 .mr(4)
16005 .nr(16)
16006 .kr(4)
16007 .sr(1)
16008 .m(4)
16009 .n(n)
16010 .k(8)
16011 .iterations(1)
16012 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16013 }
16014 }
16015
16016 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_lt_8) {
16017 TEST_REQUIRES_ARM_NEON_DOT;
16018 for (size_t k = 1; k < 8; k++) {
16019 GemmMicrokernelTester()
16020 .mr(4)
16021 .nr(16)
16022 .kr(4)
16023 .sr(1)
16024 .m(4)
16025 .n(16)
16026 .k(k)
16027 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16028 }
16029 }
16030
16031 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_lt_8_subtile) {
16032 TEST_REQUIRES_ARM_NEON_DOT;
16033 for (size_t k = 1; k < 8; k++) {
16034 for (uint32_t n = 1; n <= 16; n++) {
16035 for (uint32_t m = 1; m <= 4; m++) {
16036 GemmMicrokernelTester()
16037 .mr(4)
16038 .nr(16)
16039 .kr(4)
16040 .sr(1)
16041 .m(m)
16042 .n(n)
16043 .k(k)
16044 .iterations(1)
16045 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16046 }
16047 }
16048 }
16049 }
16050
16051 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_gt_8) {
16052 TEST_REQUIRES_ARM_NEON_DOT;
16053 for (size_t k = 9; k < 16; k++) {
16054 GemmMicrokernelTester()
16055 .mr(4)
16056 .nr(16)
16057 .kr(4)
16058 .sr(1)
16059 .m(4)
16060 .n(16)
16061 .k(k)
16062 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16063 }
16064 }
16065
16066 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_gt_8_subtile) {
16067 TEST_REQUIRES_ARM_NEON_DOT;
16068 for (size_t k = 9; k < 16; k++) {
16069 for (uint32_t n = 1; n <= 16; n++) {
16070 for (uint32_t m = 1; m <= 4; m++) {
16071 GemmMicrokernelTester()
16072 .mr(4)
16073 .nr(16)
16074 .kr(4)
16075 .sr(1)
16076 .m(m)
16077 .n(n)
16078 .k(k)
16079 .iterations(1)
16080 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16081 }
16082 }
16083 }
16084 }
16085
16086 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_div_8) {
16087 TEST_REQUIRES_ARM_NEON_DOT;
16088 for (size_t k = 16; k <= 80; k += 8) {
16089 GemmMicrokernelTester()
16090 .mr(4)
16091 .nr(16)
16092 .kr(4)
16093 .sr(1)
16094 .m(4)
16095 .n(16)
16096 .k(k)
16097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16098 }
16099 }
16100
16101 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, k_div_8_subtile) {
16102 TEST_REQUIRES_ARM_NEON_DOT;
16103 for (size_t k = 16; k <= 80; k += 8) {
16104 for (uint32_t n = 1; n <= 16; n++) {
16105 for (uint32_t m = 1; m <= 4; m++) {
16106 GemmMicrokernelTester()
16107 .mr(4)
16108 .nr(16)
16109 .kr(4)
16110 .sr(1)
16111 .m(m)
16112 .n(n)
16113 .k(k)
16114 .iterations(1)
16115 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16116 }
16117 }
16118 }
16119 }
16120
16121 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_gt_16) {
16122 TEST_REQUIRES_ARM_NEON_DOT;
16123 for (uint32_t n = 17; n < 32; n++) {
16124 for (size_t k = 1; k <= 40; k += 9) {
16125 GemmMicrokernelTester()
16126 .mr(4)
16127 .nr(16)
16128 .kr(4)
16129 .sr(1)
16130 .m(4)
16131 .n(n)
16132 .k(k)
16133 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16134 }
16135 }
16136 }
16137
16138 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_gt_16_strided_cn) {
16139 TEST_REQUIRES_ARM_NEON_DOT;
16140 for (uint32_t n = 17; n < 32; n++) {
16141 for (size_t k = 1; k <= 40; k += 9) {
16142 GemmMicrokernelTester()
16143 .mr(4)
16144 .nr(16)
16145 .kr(4)
16146 .sr(1)
16147 .m(4)
16148 .n(n)
16149 .k(k)
16150 .cn_stride(19)
16151 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16152 }
16153 }
16154 }
16155
16156 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_gt_16_subtile) {
16157 TEST_REQUIRES_ARM_NEON_DOT;
16158 for (uint32_t n = 17; n < 32; n++) {
16159 for (size_t k = 1; k <= 40; k += 9) {
16160 for (uint32_t m = 1; m <= 4; m++) {
16161 GemmMicrokernelTester()
16162 .mr(4)
16163 .nr(16)
16164 .kr(4)
16165 .sr(1)
16166 .m(m)
16167 .n(n)
16168 .k(k)
16169 .iterations(1)
16170 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16171 }
16172 }
16173 }
16174 }
16175
16176 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_div_16) {
16177 TEST_REQUIRES_ARM_NEON_DOT;
16178 for (uint32_t n = 32; n <= 48; n += 16) {
16179 for (size_t k = 1; k <= 40; k += 9) {
16180 GemmMicrokernelTester()
16181 .mr(4)
16182 .nr(16)
16183 .kr(4)
16184 .sr(1)
16185 .m(4)
16186 .n(n)
16187 .k(k)
16188 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16189 }
16190 }
16191 }
16192
16193 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_div_16_strided_cn) {
16194 TEST_REQUIRES_ARM_NEON_DOT;
16195 for (uint32_t n = 32; n <= 48; n += 16) {
16196 for (size_t k = 1; k <= 40; k += 9) {
16197 GemmMicrokernelTester()
16198 .mr(4)
16199 .nr(16)
16200 .kr(4)
16201 .sr(1)
16202 .m(4)
16203 .n(n)
16204 .k(k)
16205 .cn_stride(19)
16206 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16207 }
16208 }
16209 }
16210
16211 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_div_16_subtile) {
16212 TEST_REQUIRES_ARM_NEON_DOT;
16213 for (uint32_t n = 32; n <= 48; n += 16) {
16214 for (size_t k = 1; k <= 40; k += 9) {
16215 for (uint32_t m = 1; m <= 4; m++) {
16216 GemmMicrokernelTester()
16217 .mr(4)
16218 .nr(16)
16219 .kr(4)
16220 .sr(1)
16221 .m(m)
16222 .n(n)
16223 .k(k)
16224 .iterations(1)
16225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16226 }
16227 }
16228 }
16229 }
16230
16231 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, small_kernel) {
16232 TEST_REQUIRES_ARM_NEON_DOT;
16233 for (size_t k = 1; k <= 40; k += 9) {
16234 GemmMicrokernelTester()
16235 .mr(4)
16236 .nr(16)
16237 .kr(4)
16238 .sr(1)
16239 .m(4)
16240 .n(16)
16241 .k(k)
16242 .ks(3)
16243 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16244 }
16245 }
16246
16247 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, small_kernel_subtile) {
16248 TEST_REQUIRES_ARM_NEON_DOT;
16249 for (size_t k = 1; k <= 40; k += 9) {
16250 for (uint32_t n = 1; n <= 16; n++) {
16251 for (uint32_t m = 1; m <= 4; m++) {
16252 GemmMicrokernelTester()
16253 .mr(4)
16254 .nr(16)
16255 .kr(4)
16256 .sr(1)
16257 .m(m)
16258 .n(n)
16259 .k(k)
16260 .ks(3)
16261 .iterations(1)
16262 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16263 }
16264 }
16265 }
16266 }
16267
16268 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_gt_16_small_kernel) {
16269 TEST_REQUIRES_ARM_NEON_DOT;
16270 for (uint32_t n = 17; n < 32; n++) {
16271 for (size_t k = 1; k <= 40; k += 9) {
16272 GemmMicrokernelTester()
16273 .mr(4)
16274 .nr(16)
16275 .kr(4)
16276 .sr(1)
16277 .m(4)
16278 .n(n)
16279 .k(k)
16280 .ks(3)
16281 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16282 }
16283 }
16284 }
16285
16286 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, n_div_16_small_kernel) {
16287 TEST_REQUIRES_ARM_NEON_DOT;
16288 for (uint32_t n = 32; n <= 48; n += 16) {
16289 for (size_t k = 1; k <= 40; k += 9) {
16290 GemmMicrokernelTester()
16291 .mr(4)
16292 .nr(16)
16293 .kr(4)
16294 .sr(1)
16295 .m(4)
16296 .n(n)
16297 .k(k)
16298 .ks(3)
16299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16300 }
16301 }
16302 }
16303
16304 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, strided_cm_subtile) {
16305 TEST_REQUIRES_ARM_NEON_DOT;
16306 for (size_t k = 1; k <= 40; k += 9) {
16307 for (uint32_t n = 1; n <= 16; n++) {
16308 for (uint32_t m = 1; m <= 4; m++) {
16309 GemmMicrokernelTester()
16310 .mr(4)
16311 .nr(16)
16312 .kr(4)
16313 .sr(1)
16314 .m(m)
16315 .n(n)
16316 .k(k)
16317 .cm_stride(19)
16318 .iterations(1)
16319 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16320 }
16321 }
16322 }
16323 }
16324
16325 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, a_offset) {
16326 TEST_REQUIRES_ARM_NEON_DOT;
16327 for (size_t k = 1; k <= 40; k += 9) {
16328 GemmMicrokernelTester()
16329 .mr(4)
16330 .nr(16)
16331 .kr(4)
16332 .sr(1)
16333 .m(4)
16334 .n(16)
16335 .k(k)
16336 .ks(3)
16337 .a_offset(163)
16338 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16339 }
16340 }
16341
16342 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, zero) {
16343 TEST_REQUIRES_ARM_NEON_DOT;
16344 for (size_t k = 1; k <= 40; k += 9) {
16345 for (uint32_t mz = 0; mz < 4; mz++) {
16346 GemmMicrokernelTester()
16347 .mr(4)
16348 .nr(16)
16349 .kr(4)
16350 .sr(1)
16351 .m(4)
16352 .n(16)
16353 .k(k)
16354 .ks(3)
16355 .a_offset(163)
16356 .zero_index(mz)
16357 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16358 }
16359 }
16360 }
16361
16362 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, qmin) {
16363 TEST_REQUIRES_ARM_NEON_DOT;
16364 GemmMicrokernelTester()
16365 .mr(4)
16366 .nr(16)
16367 .kr(4)
16368 .sr(1)
16369 .m(4)
16370 .n(16)
16371 .k(8)
16372 .qmin(128)
16373 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16374 }
16375
16376 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, qmax) {
16377 TEST_REQUIRES_ARM_NEON_DOT;
16378 GemmMicrokernelTester()
16379 .mr(4)
16380 .nr(16)
16381 .kr(4)
16382 .sr(1)
16383 .m(4)
16384 .n(16)
16385 .k(8)
16386 .qmax(128)
16387 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16388 }
16389
16390 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD64, strided_cm) {
16391 TEST_REQUIRES_ARM_NEON_DOT;
16392 GemmMicrokernelTester()
16393 .mr(4)
16394 .nr(16)
16395 .kr(4)
16396 .sr(1)
16397 .m(4)
16398 .n(16)
16399 .k(8)
16400 .cm_stride(19)
16401 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16402 }
16403#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
16404
16405
16406#if XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
16407 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16) {
16408 TEST_REQUIRES_ARM_NEON_DOT;
16409 GemmMicrokernelTester()
16410 .mr(4)
16411 .nr(16)
16412 .kr(4)
16413 .sr(1)
16414 .m(4)
16415 .n(16)
16416 .k(16)
16417 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16418 }
16419
16420 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, strided_cn) {
16421 TEST_REQUIRES_ARM_NEON_DOT;
16422 GemmMicrokernelTester()
16423 .mr(4)
16424 .nr(16)
16425 .kr(4)
16426 .sr(1)
16427 .m(4)
16428 .n(16)
16429 .k(16)
16430 .cn_stride(19)
16431 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16432 }
16433
16434 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile) {
16435 TEST_REQUIRES_ARM_NEON_DOT;
16436 for (uint32_t n = 1; n <= 16; n++) {
16437 for (uint32_t m = 1; m <= 4; m++) {
16438 GemmMicrokernelTester()
16439 .mr(4)
16440 .nr(16)
16441 .kr(4)
16442 .sr(1)
16443 .m(m)
16444 .n(n)
16445 .k(16)
16446 .iterations(1)
16447 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16448 }
16449 }
16450 }
16451
16452 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile_m) {
16453 TEST_REQUIRES_ARM_NEON_DOT;
16454 for (uint32_t m = 1; m <= 4; m++) {
16455 GemmMicrokernelTester()
16456 .mr(4)
16457 .nr(16)
16458 .kr(4)
16459 .sr(1)
16460 .m(m)
16461 .n(16)
16462 .k(16)
16463 .iterations(1)
16464 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16465 }
16466 }
16467
16468 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_eq_16_subtile_n) {
16469 TEST_REQUIRES_ARM_NEON_DOT;
16470 for (uint32_t n = 1; n <= 16; n++) {
16471 GemmMicrokernelTester()
16472 .mr(4)
16473 .nr(16)
16474 .kr(4)
16475 .sr(1)
16476 .m(4)
16477 .n(n)
16478 .k(16)
16479 .iterations(1)
16480 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16481 }
16482 }
16483
16484 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_lt_16) {
16485 TEST_REQUIRES_ARM_NEON_DOT;
16486 for (size_t k = 1; k < 16; k++) {
16487 GemmMicrokernelTester()
16488 .mr(4)
16489 .nr(16)
16490 .kr(4)
16491 .sr(1)
16492 .m(4)
16493 .n(16)
16494 .k(k)
16495 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16496 }
16497 }
16498
16499 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_lt_16_subtile) {
16500 TEST_REQUIRES_ARM_NEON_DOT;
16501 for (size_t k = 1; k < 16; k++) {
16502 for (uint32_t n = 1; n <= 16; n++) {
16503 for (uint32_t m = 1; m <= 4; m++) {
16504 GemmMicrokernelTester()
16505 .mr(4)
16506 .nr(16)
16507 .kr(4)
16508 .sr(1)
16509 .m(m)
16510 .n(n)
16511 .k(k)
16512 .iterations(1)
16513 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16514 }
16515 }
16516 }
16517 }
16518
16519 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_gt_16) {
16520 TEST_REQUIRES_ARM_NEON_DOT;
16521 for (size_t k = 17; k < 32; k++) {
16522 GemmMicrokernelTester()
16523 .mr(4)
16524 .nr(16)
16525 .kr(4)
16526 .sr(1)
16527 .m(4)
16528 .n(16)
16529 .k(k)
16530 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16531 }
16532 }
16533
16534 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_gt_16_subtile) {
16535 TEST_REQUIRES_ARM_NEON_DOT;
16536 for (size_t k = 17; k < 32; k++) {
16537 for (uint32_t n = 1; n <= 16; n++) {
16538 for (uint32_t m = 1; m <= 4; m++) {
16539 GemmMicrokernelTester()
16540 .mr(4)
16541 .nr(16)
16542 .kr(4)
16543 .sr(1)
16544 .m(m)
16545 .n(n)
16546 .k(k)
16547 .iterations(1)
16548 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16549 }
16550 }
16551 }
16552 }
16553
16554 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_div_16) {
16555 TEST_REQUIRES_ARM_NEON_DOT;
16556 for (size_t k = 32; k <= 160; k += 16) {
16557 GemmMicrokernelTester()
16558 .mr(4)
16559 .nr(16)
16560 .kr(4)
16561 .sr(1)
16562 .m(4)
16563 .n(16)
16564 .k(k)
16565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16566 }
16567 }
16568
16569 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, k_div_16_subtile) {
16570 TEST_REQUIRES_ARM_NEON_DOT;
16571 for (size_t k = 32; k <= 160; k += 16) {
16572 for (uint32_t n = 1; n <= 16; n++) {
16573 for (uint32_t m = 1; m <= 4; m++) {
16574 GemmMicrokernelTester()
16575 .mr(4)
16576 .nr(16)
16577 .kr(4)
16578 .sr(1)
16579 .m(m)
16580 .n(n)
16581 .k(k)
16582 .iterations(1)
16583 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16584 }
16585 }
16586 }
16587 }
16588
16589 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16) {
16590 TEST_REQUIRES_ARM_NEON_DOT;
16591 for (uint32_t n = 17; n < 32; n++) {
16592 for (size_t k = 1; k <= 80; k += 17) {
16593 GemmMicrokernelTester()
16594 .mr(4)
16595 .nr(16)
16596 .kr(4)
16597 .sr(1)
16598 .m(4)
16599 .n(n)
16600 .k(k)
16601 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16602 }
16603 }
16604 }
16605
16606 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16_strided_cn) {
16607 TEST_REQUIRES_ARM_NEON_DOT;
16608 for (uint32_t n = 17; n < 32; n++) {
16609 for (size_t k = 1; k <= 80; k += 17) {
16610 GemmMicrokernelTester()
16611 .mr(4)
16612 .nr(16)
16613 .kr(4)
16614 .sr(1)
16615 .m(4)
16616 .n(n)
16617 .k(k)
16618 .cn_stride(19)
16619 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16620 }
16621 }
16622 }
16623
16624 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16_subtile) {
16625 TEST_REQUIRES_ARM_NEON_DOT;
16626 for (uint32_t n = 17; n < 32; n++) {
16627 for (size_t k = 1; k <= 80; k += 17) {
16628 for (uint32_t m = 1; m <= 4; m++) {
16629 GemmMicrokernelTester()
16630 .mr(4)
16631 .nr(16)
16632 .kr(4)
16633 .sr(1)
16634 .m(m)
16635 .n(n)
16636 .k(k)
16637 .iterations(1)
16638 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16639 }
16640 }
16641 }
16642 }
16643
16644 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16) {
16645 TEST_REQUIRES_ARM_NEON_DOT;
16646 for (uint32_t n = 32; n <= 48; n += 16) {
16647 for (size_t k = 1; k <= 80; k += 17) {
16648 GemmMicrokernelTester()
16649 .mr(4)
16650 .nr(16)
16651 .kr(4)
16652 .sr(1)
16653 .m(4)
16654 .n(n)
16655 .k(k)
16656 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16657 }
16658 }
16659 }
16660
16661 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16_strided_cn) {
16662 TEST_REQUIRES_ARM_NEON_DOT;
16663 for (uint32_t n = 32; n <= 48; n += 16) {
16664 for (size_t k = 1; k <= 80; k += 17) {
16665 GemmMicrokernelTester()
16666 .mr(4)
16667 .nr(16)
16668 .kr(4)
16669 .sr(1)
16670 .m(4)
16671 .n(n)
16672 .k(k)
16673 .cn_stride(19)
16674 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16675 }
16676 }
16677 }
16678
16679 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16_subtile) {
16680 TEST_REQUIRES_ARM_NEON_DOT;
16681 for (uint32_t n = 32; n <= 48; n += 16) {
16682 for (size_t k = 1; k <= 80; k += 17) {
16683 for (uint32_t m = 1; m <= 4; m++) {
16684 GemmMicrokernelTester()
16685 .mr(4)
16686 .nr(16)
16687 .kr(4)
16688 .sr(1)
16689 .m(m)
16690 .n(n)
16691 .k(k)
16692 .iterations(1)
16693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16694 }
16695 }
16696 }
16697 }
16698
16699 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, small_kernel) {
16700 TEST_REQUIRES_ARM_NEON_DOT;
16701 for (size_t k = 1; k <= 80; k += 17) {
16702 GemmMicrokernelTester()
16703 .mr(4)
16704 .nr(16)
16705 .kr(4)
16706 .sr(1)
16707 .m(4)
16708 .n(16)
16709 .k(k)
16710 .ks(3)
16711 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16712 }
16713 }
16714
16715 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, small_kernel_subtile) {
16716 TEST_REQUIRES_ARM_NEON_DOT;
16717 for (size_t k = 1; k <= 80; k += 17) {
16718 for (uint32_t n = 1; n <= 16; n++) {
16719 for (uint32_t m = 1; m <= 4; m++) {
16720 GemmMicrokernelTester()
16721 .mr(4)
16722 .nr(16)
16723 .kr(4)
16724 .sr(1)
16725 .m(m)
16726 .n(n)
16727 .k(k)
16728 .ks(3)
16729 .iterations(1)
16730 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16731 }
16732 }
16733 }
16734 }
16735
16736 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_gt_16_small_kernel) {
16737 TEST_REQUIRES_ARM_NEON_DOT;
16738 for (uint32_t n = 17; n < 32; n++) {
16739 for (size_t k = 1; k <= 80; k += 17) {
16740 GemmMicrokernelTester()
16741 .mr(4)
16742 .nr(16)
16743 .kr(4)
16744 .sr(1)
16745 .m(4)
16746 .n(n)
16747 .k(k)
16748 .ks(3)
16749 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16750 }
16751 }
16752 }
16753
16754 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, n_div_16_small_kernel) {
16755 TEST_REQUIRES_ARM_NEON_DOT;
16756 for (uint32_t n = 32; n <= 48; n += 16) {
16757 for (size_t k = 1; k <= 80; k += 17) {
16758 GemmMicrokernelTester()
16759 .mr(4)
16760 .nr(16)
16761 .kr(4)
16762 .sr(1)
16763 .m(4)
16764 .n(n)
16765 .k(k)
16766 .ks(3)
16767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16768 }
16769 }
16770 }
16771
16772 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, strided_cm_subtile) {
16773 TEST_REQUIRES_ARM_NEON_DOT;
16774 for (size_t k = 1; k <= 80; k += 17) {
16775 for (uint32_t n = 1; n <= 16; n++) {
16776 for (uint32_t m = 1; m <= 4; m++) {
16777 GemmMicrokernelTester()
16778 .mr(4)
16779 .nr(16)
16780 .kr(4)
16781 .sr(1)
16782 .m(m)
16783 .n(n)
16784 .k(k)
16785 .cm_stride(19)
16786 .iterations(1)
16787 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16788 }
16789 }
16790 }
16791 }
16792
16793 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, a_offset) {
16794 TEST_REQUIRES_ARM_NEON_DOT;
16795 for (size_t k = 1; k <= 80; k += 17) {
16796 GemmMicrokernelTester()
16797 .mr(4)
16798 .nr(16)
16799 .kr(4)
16800 .sr(1)
16801 .m(4)
16802 .n(16)
16803 .k(k)
16804 .ks(3)
16805 .a_offset(331)
16806 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16807 }
16808 }
16809
16810 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, zero) {
16811 TEST_REQUIRES_ARM_NEON_DOT;
16812 for (size_t k = 1; k <= 80; k += 17) {
16813 for (uint32_t mz = 0; mz < 4; mz++) {
16814 GemmMicrokernelTester()
16815 .mr(4)
16816 .nr(16)
16817 .kr(4)
16818 .sr(1)
16819 .m(4)
16820 .n(16)
16821 .k(k)
16822 .ks(3)
16823 .a_offset(331)
16824 .zero_index(mz)
16825 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16826 }
16827 }
16828 }
16829
16830 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, qmin) {
16831 TEST_REQUIRES_ARM_NEON_DOT;
16832 GemmMicrokernelTester()
16833 .mr(4)
16834 .nr(16)
16835 .kr(4)
16836 .sr(1)
16837 .m(4)
16838 .n(16)
16839 .k(16)
16840 .qmin(128)
16841 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16842 }
16843
16844 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, qmax) {
16845 TEST_REQUIRES_ARM_NEON_DOT;
16846 GemmMicrokernelTester()
16847 .mr(4)
16848 .nr(16)
16849 .kr(4)
16850 .sr(1)
16851 .m(4)
16852 .n(16)
16853 .k(16)
16854 .qmax(128)
16855 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16856 }
16857
16858 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C4__AARCH64_NEONDOT_LD128, strided_cm) {
16859 TEST_REQUIRES_ARM_NEON_DOT;
16860 GemmMicrokernelTester()
16861 .mr(4)
16862 .nr(16)
16863 .kr(4)
16864 .sr(1)
16865 .m(4)
16866 .n(16)
16867 .k(16)
16868 .cm_stride(19)
16869 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c4__aarch64_neondot_ld128, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16870 }
16871#endif // XNN_ARCH_ARM64 && XNN_ENABLE_ASSEMBLY
16872
16873
16874#if XNN_ARCH_ARM || XNN_ARCH_ARM64
16875 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8) {
16876 TEST_REQUIRES_ARM_NEON;
16877 GemmMicrokernelTester()
16878 .mr(3)
16879 .nr(8)
16880 .kr(8)
16881 .sr(1)
16882 .m(3)
16883 .n(8)
16884 .k(8)
16885 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16886 }
16887
16888 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cn) {
16889 TEST_REQUIRES_ARM_NEON;
16890 GemmMicrokernelTester()
16891 .mr(3)
16892 .nr(8)
16893 .kr(8)
16894 .sr(1)
16895 .m(3)
16896 .n(8)
16897 .k(8)
16898 .cn_stride(11)
16899 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16900 }
16901
16902 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile) {
16903 TEST_REQUIRES_ARM_NEON;
16904 for (uint32_t n = 1; n <= 8; n++) {
16905 for (uint32_t m = 1; m <= 3; m++) {
16906 GemmMicrokernelTester()
16907 .mr(3)
16908 .nr(8)
16909 .kr(8)
16910 .sr(1)
16911 .m(m)
16912 .n(n)
16913 .k(8)
16914 .iterations(1)
16915 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16916 }
16917 }
16918 }
16919
16920 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile_m) {
16921 TEST_REQUIRES_ARM_NEON;
16922 for (uint32_t m = 1; m <= 3; m++) {
16923 GemmMicrokernelTester()
16924 .mr(3)
16925 .nr(8)
16926 .kr(8)
16927 .sr(1)
16928 .m(m)
16929 .n(8)
16930 .k(8)
16931 .iterations(1)
16932 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16933 }
16934 }
16935
16936 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_eq_8_subtile_n) {
16937 TEST_REQUIRES_ARM_NEON;
16938 for (uint32_t n = 1; n <= 8; n++) {
16939 GemmMicrokernelTester()
16940 .mr(3)
16941 .nr(8)
16942 .kr(8)
16943 .sr(1)
16944 .m(3)
16945 .n(n)
16946 .k(8)
16947 .iterations(1)
16948 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16949 }
16950 }
16951
16952 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_lt_8) {
16953 TEST_REQUIRES_ARM_NEON;
16954 for (size_t k = 1; k < 8; k++) {
16955 GemmMicrokernelTester()
16956 .mr(3)
16957 .nr(8)
16958 .kr(8)
16959 .sr(1)
16960 .m(3)
16961 .n(8)
16962 .k(k)
16963 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16964 }
16965 }
16966
16967 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_lt_8_subtile) {
16968 TEST_REQUIRES_ARM_NEON;
16969 for (size_t k = 1; k < 8; k++) {
16970 for (uint32_t n = 1; n <= 8; n++) {
16971 for (uint32_t m = 1; m <= 3; m++) {
16972 GemmMicrokernelTester()
16973 .mr(3)
16974 .nr(8)
16975 .kr(8)
16976 .sr(1)
16977 .m(m)
16978 .n(n)
16979 .k(k)
16980 .iterations(1)
16981 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16982 }
16983 }
16984 }
16985 }
16986
16987 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_gt_8) {
16988 TEST_REQUIRES_ARM_NEON;
16989 for (size_t k = 9; k < 16; k++) {
16990 GemmMicrokernelTester()
16991 .mr(3)
16992 .nr(8)
16993 .kr(8)
16994 .sr(1)
16995 .m(3)
16996 .n(8)
16997 .k(k)
16998 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
16999 }
17000 }
17001
17002 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_gt_8_subtile) {
17003 TEST_REQUIRES_ARM_NEON;
17004 for (size_t k = 9; k < 16; k++) {
17005 for (uint32_t n = 1; n <= 8; n++) {
17006 for (uint32_t m = 1; m <= 3; m++) {
17007 GemmMicrokernelTester()
17008 .mr(3)
17009 .nr(8)
17010 .kr(8)
17011 .sr(1)
17012 .m(m)
17013 .n(n)
17014 .k(k)
17015 .iterations(1)
17016 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17017 }
17018 }
17019 }
17020 }
17021
17022 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_div_8) {
17023 TEST_REQUIRES_ARM_NEON;
17024 for (size_t k = 16; k <= 80; k += 8) {
17025 GemmMicrokernelTester()
17026 .mr(3)
17027 .nr(8)
17028 .kr(8)
17029 .sr(1)
17030 .m(3)
17031 .n(8)
17032 .k(k)
17033 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17034 }
17035 }
17036
17037 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, k_div_8_subtile) {
17038 TEST_REQUIRES_ARM_NEON;
17039 for (size_t k = 16; k <= 80; k += 8) {
17040 for (uint32_t n = 1; n <= 8; n++) {
17041 for (uint32_t m = 1; m <= 3; m++) {
17042 GemmMicrokernelTester()
17043 .mr(3)
17044 .nr(8)
17045 .kr(8)
17046 .sr(1)
17047 .m(m)
17048 .n(n)
17049 .k(k)
17050 .iterations(1)
17051 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17052 }
17053 }
17054 }
17055 }
17056
17057 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8) {
17058 TEST_REQUIRES_ARM_NEON;
17059 for (uint32_t n = 9; n < 16; n++) {
17060 for (size_t k = 1; k <= 40; k += 9) {
17061 GemmMicrokernelTester()
17062 .mr(3)
17063 .nr(8)
17064 .kr(8)
17065 .sr(1)
17066 .m(3)
17067 .n(n)
17068 .k(k)
17069 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17070 }
17071 }
17072 }
17073
17074 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_strided_cn) {
17075 TEST_REQUIRES_ARM_NEON;
17076 for (uint32_t n = 9; n < 16; n++) {
17077 for (size_t k = 1; k <= 40; k += 9) {
17078 GemmMicrokernelTester()
17079 .mr(3)
17080 .nr(8)
17081 .kr(8)
17082 .sr(1)
17083 .m(3)
17084 .n(n)
17085 .k(k)
17086 .cn_stride(11)
17087 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17088 }
17089 }
17090 }
17091
17092 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_subtile) {
17093 TEST_REQUIRES_ARM_NEON;
17094 for (uint32_t n = 9; n < 16; n++) {
17095 for (size_t k = 1; k <= 40; k += 9) {
17096 for (uint32_t m = 1; m <= 3; m++) {
17097 GemmMicrokernelTester()
17098 .mr(3)
17099 .nr(8)
17100 .kr(8)
17101 .sr(1)
17102 .m(m)
17103 .n(n)
17104 .k(k)
17105 .iterations(1)
17106 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17107 }
17108 }
17109 }
17110 }
17111
17112 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8) {
17113 TEST_REQUIRES_ARM_NEON;
17114 for (uint32_t n = 16; n <= 24; n += 8) {
17115 for (size_t k = 1; k <= 40; k += 9) {
17116 GemmMicrokernelTester()
17117 .mr(3)
17118 .nr(8)
17119 .kr(8)
17120 .sr(1)
17121 .m(3)
17122 .n(n)
17123 .k(k)
17124 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17125 }
17126 }
17127 }
17128
17129 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_strided_cn) {
17130 TEST_REQUIRES_ARM_NEON;
17131 for (uint32_t n = 16; n <= 24; n += 8) {
17132 for (size_t k = 1; k <= 40; k += 9) {
17133 GemmMicrokernelTester()
17134 .mr(3)
17135 .nr(8)
17136 .kr(8)
17137 .sr(1)
17138 .m(3)
17139 .n(n)
17140 .k(k)
17141 .cn_stride(11)
17142 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17143 }
17144 }
17145 }
17146
17147 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_subtile) {
17148 TEST_REQUIRES_ARM_NEON;
17149 for (uint32_t n = 16; n <= 24; n += 8) {
17150 for (size_t k = 1; k <= 40; k += 9) {
17151 for (uint32_t m = 1; m <= 3; m++) {
17152 GemmMicrokernelTester()
17153 .mr(3)
17154 .nr(8)
17155 .kr(8)
17156 .sr(1)
17157 .m(m)
17158 .n(n)
17159 .k(k)
17160 .iterations(1)
17161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17162 }
17163 }
17164 }
17165 }
17166
17167 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, small_kernel) {
17168 TEST_REQUIRES_ARM_NEON;
17169 for (size_t k = 1; k <= 40; k += 9) {
17170 GemmMicrokernelTester()
17171 .mr(3)
17172 .nr(8)
17173 .kr(8)
17174 .sr(1)
17175 .m(3)
17176 .n(8)
17177 .k(k)
17178 .ks(3)
17179 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17180 }
17181 }
17182
17183 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, small_kernel_subtile) {
17184 TEST_REQUIRES_ARM_NEON;
17185 for (size_t k = 1; k <= 40; k += 9) {
17186 for (uint32_t n = 1; n <= 8; n++) {
17187 for (uint32_t m = 1; m <= 3; m++) {
17188 GemmMicrokernelTester()
17189 .mr(3)
17190 .nr(8)
17191 .kr(8)
17192 .sr(1)
17193 .m(m)
17194 .n(n)
17195 .k(k)
17196 .ks(3)
17197 .iterations(1)
17198 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17199 }
17200 }
17201 }
17202 }
17203
17204 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_gt_8_small_kernel) {
17205 TEST_REQUIRES_ARM_NEON;
17206 for (uint32_t n = 9; n < 16; n++) {
17207 for (size_t k = 1; k <= 40; k += 9) {
17208 GemmMicrokernelTester()
17209 .mr(3)
17210 .nr(8)
17211 .kr(8)
17212 .sr(1)
17213 .m(3)
17214 .n(n)
17215 .k(k)
17216 .ks(3)
17217 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17218 }
17219 }
17220 }
17221
17222 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, n_div_8_small_kernel) {
17223 TEST_REQUIRES_ARM_NEON;
17224 for (uint32_t n = 16; n <= 24; n += 8) {
17225 for (size_t k = 1; k <= 40; k += 9) {
17226 GemmMicrokernelTester()
17227 .mr(3)
17228 .nr(8)
17229 .kr(8)
17230 .sr(1)
17231 .m(3)
17232 .n(n)
17233 .k(k)
17234 .ks(3)
17235 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17236 }
17237 }
17238 }
17239
17240 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cm_subtile) {
17241 TEST_REQUIRES_ARM_NEON;
17242 for (size_t k = 1; k <= 40; k += 9) {
17243 for (uint32_t n = 1; n <= 8; n++) {
17244 for (uint32_t m = 1; m <= 3; m++) {
17245 GemmMicrokernelTester()
17246 .mr(3)
17247 .nr(8)
17248 .kr(8)
17249 .sr(1)
17250 .m(m)
17251 .n(n)
17252 .k(k)
17253 .cm_stride(11)
17254 .iterations(1)
17255 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17256 }
17257 }
17258 }
17259 }
17260
17261 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, a_offset) {
17262 TEST_REQUIRES_ARM_NEON;
17263 for (size_t k = 1; k <= 40; k += 9) {
17264 GemmMicrokernelTester()
17265 .mr(3)
17266 .nr(8)
17267 .kr(8)
17268 .sr(1)
17269 .m(3)
17270 .n(8)
17271 .k(k)
17272 .ks(3)
17273 .a_offset(127)
17274 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17275 }
17276 }
17277
17278 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, zero) {
17279 TEST_REQUIRES_ARM_NEON;
17280 for (size_t k = 1; k <= 40; k += 9) {
17281 for (uint32_t mz = 0; mz < 3; mz++) {
17282 GemmMicrokernelTester()
17283 .mr(3)
17284 .nr(8)
17285 .kr(8)
17286 .sr(1)
17287 .m(3)
17288 .n(8)
17289 .k(k)
17290 .ks(3)
17291 .a_offset(127)
17292 .zero_index(mz)
17293 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17294 }
17295 }
17296 }
17297
17298 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, qmin) {
17299 TEST_REQUIRES_ARM_NEON;
17300 GemmMicrokernelTester()
17301 .mr(3)
17302 .nr(8)
17303 .kr(8)
17304 .sr(1)
17305 .m(3)
17306 .n(8)
17307 .k(8)
17308 .qmin(128)
17309 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17310 }
17311
17312 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, qmax) {
17313 TEST_REQUIRES_ARM_NEON;
17314 GemmMicrokernelTester()
17315 .mr(3)
17316 .nr(8)
17317 .kr(8)
17318 .sr(1)
17319 .m(3)
17320 .n(8)
17321 .k(8)
17322 .qmax(128)
17323 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17324 }
17325
17326 TEST(QS8_IGEMM_MINMAX_RNDNU_3X8C8__NEON_MULL, strided_cm) {
17327 TEST_REQUIRES_ARM_NEON;
17328 GemmMicrokernelTester()
17329 .mr(3)
17330 .nr(8)
17331 .kr(8)
17332 .sr(1)
17333 .m(3)
17334 .n(8)
17335 .k(8)
17336 .cm_stride(11)
17337 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x8c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17338 }
17339#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
17340
17341
17342#if XNN_ARCH_ARM || XNN_ARCH_ARM64
17343 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8) {
17344 TEST_REQUIRES_ARM_NEON;
17345 GemmMicrokernelTester()
17346 .mr(1)
17347 .nr(16)
17348 .kr(8)
17349 .sr(1)
17350 .m(1)
17351 .n(16)
17352 .k(8)
17353 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17354 }
17355
17356 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, strided_cn) {
17357 TEST_REQUIRES_ARM_NEON;
17358 GemmMicrokernelTester()
17359 .mr(1)
17360 .nr(16)
17361 .kr(8)
17362 .sr(1)
17363 .m(1)
17364 .n(16)
17365 .k(8)
17366 .cn_stride(19)
17367 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17368 }
17369
17370 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8_subtile) {
17371 TEST_REQUIRES_ARM_NEON;
17372 for (uint32_t n = 1; n <= 16; n++) {
17373 for (uint32_t m = 1; m <= 1; m++) {
17374 GemmMicrokernelTester()
17375 .mr(1)
17376 .nr(16)
17377 .kr(8)
17378 .sr(1)
17379 .m(m)
17380 .n(n)
17381 .k(8)
17382 .iterations(1)
17383 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17384 }
17385 }
17386 }
17387
17388 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8_subtile_m) {
17389 TEST_REQUIRES_ARM_NEON;
17390 for (uint32_t m = 1; m <= 1; m++) {
17391 GemmMicrokernelTester()
17392 .mr(1)
17393 .nr(16)
17394 .kr(8)
17395 .sr(1)
17396 .m(m)
17397 .n(16)
17398 .k(8)
17399 .iterations(1)
17400 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17401 }
17402 }
17403
17404 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_eq_8_subtile_n) {
17405 TEST_REQUIRES_ARM_NEON;
17406 for (uint32_t n = 1; n <= 16; n++) {
17407 GemmMicrokernelTester()
17408 .mr(1)
17409 .nr(16)
17410 .kr(8)
17411 .sr(1)
17412 .m(1)
17413 .n(n)
17414 .k(8)
17415 .iterations(1)
17416 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17417 }
17418 }
17419
17420 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_lt_8) {
17421 TEST_REQUIRES_ARM_NEON;
17422 for (size_t k = 1; k < 8; k++) {
17423 GemmMicrokernelTester()
17424 .mr(1)
17425 .nr(16)
17426 .kr(8)
17427 .sr(1)
17428 .m(1)
17429 .n(16)
17430 .k(k)
17431 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17432 }
17433 }
17434
17435 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_lt_8_subtile) {
17436 TEST_REQUIRES_ARM_NEON;
17437 for (size_t k = 1; k < 8; k++) {
17438 for (uint32_t n = 1; n <= 16; n++) {
17439 for (uint32_t m = 1; m <= 1; m++) {
17440 GemmMicrokernelTester()
17441 .mr(1)
17442 .nr(16)
17443 .kr(8)
17444 .sr(1)
17445 .m(m)
17446 .n(n)
17447 .k(k)
17448 .iterations(1)
17449 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17450 }
17451 }
17452 }
17453 }
17454
17455 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_gt_8) {
17456 TEST_REQUIRES_ARM_NEON;
17457 for (size_t k = 9; k < 16; k++) {
17458 GemmMicrokernelTester()
17459 .mr(1)
17460 .nr(16)
17461 .kr(8)
17462 .sr(1)
17463 .m(1)
17464 .n(16)
17465 .k(k)
17466 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17467 }
17468 }
17469
17470 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_gt_8_subtile) {
17471 TEST_REQUIRES_ARM_NEON;
17472 for (size_t k = 9; k < 16; k++) {
17473 for (uint32_t n = 1; n <= 16; n++) {
17474 for (uint32_t m = 1; m <= 1; m++) {
17475 GemmMicrokernelTester()
17476 .mr(1)
17477 .nr(16)
17478 .kr(8)
17479 .sr(1)
17480 .m(m)
17481 .n(n)
17482 .k(k)
17483 .iterations(1)
17484 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17485 }
17486 }
17487 }
17488 }
17489
17490 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_div_8) {
17491 TEST_REQUIRES_ARM_NEON;
17492 for (size_t k = 16; k <= 80; k += 8) {
17493 GemmMicrokernelTester()
17494 .mr(1)
17495 .nr(16)
17496 .kr(8)
17497 .sr(1)
17498 .m(1)
17499 .n(16)
17500 .k(k)
17501 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17502 }
17503 }
17504
17505 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, k_div_8_subtile) {
17506 TEST_REQUIRES_ARM_NEON;
17507 for (size_t k = 16; k <= 80; k += 8) {
17508 for (uint32_t n = 1; n <= 16; n++) {
17509 for (uint32_t m = 1; m <= 1; m++) {
17510 GemmMicrokernelTester()
17511 .mr(1)
17512 .nr(16)
17513 .kr(8)
17514 .sr(1)
17515 .m(m)
17516 .n(n)
17517 .k(k)
17518 .iterations(1)
17519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17520 }
17521 }
17522 }
17523 }
17524
17525 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16) {
17526 TEST_REQUIRES_ARM_NEON;
17527 for (uint32_t n = 17; n < 32; n++) {
17528 for (size_t k = 1; k <= 40; k += 9) {
17529 GemmMicrokernelTester()
17530 .mr(1)
17531 .nr(16)
17532 .kr(8)
17533 .sr(1)
17534 .m(1)
17535 .n(n)
17536 .k(k)
17537 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17538 }
17539 }
17540 }
17541
17542 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16_strided_cn) {
17543 TEST_REQUIRES_ARM_NEON;
17544 for (uint32_t n = 17; n < 32; n++) {
17545 for (size_t k = 1; k <= 40; k += 9) {
17546 GemmMicrokernelTester()
17547 .mr(1)
17548 .nr(16)
17549 .kr(8)
17550 .sr(1)
17551 .m(1)
17552 .n(n)
17553 .k(k)
17554 .cn_stride(19)
17555 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17556 }
17557 }
17558 }
17559
17560 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16_subtile) {
17561 TEST_REQUIRES_ARM_NEON;
17562 for (uint32_t n = 17; n < 32; n++) {
17563 for (size_t k = 1; k <= 40; k += 9) {
17564 for (uint32_t m = 1; m <= 1; m++) {
17565 GemmMicrokernelTester()
17566 .mr(1)
17567 .nr(16)
17568 .kr(8)
17569 .sr(1)
17570 .m(m)
17571 .n(n)
17572 .k(k)
17573 .iterations(1)
17574 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17575 }
17576 }
17577 }
17578 }
17579
17580 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16) {
17581 TEST_REQUIRES_ARM_NEON;
17582 for (uint32_t n = 32; n <= 48; n += 16) {
17583 for (size_t k = 1; k <= 40; k += 9) {
17584 GemmMicrokernelTester()
17585 .mr(1)
17586 .nr(16)
17587 .kr(8)
17588 .sr(1)
17589 .m(1)
17590 .n(n)
17591 .k(k)
17592 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17593 }
17594 }
17595 }
17596
17597 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16_strided_cn) {
17598 TEST_REQUIRES_ARM_NEON;
17599 for (uint32_t n = 32; n <= 48; n += 16) {
17600 for (size_t k = 1; k <= 40; k += 9) {
17601 GemmMicrokernelTester()
17602 .mr(1)
17603 .nr(16)
17604 .kr(8)
17605 .sr(1)
17606 .m(1)
17607 .n(n)
17608 .k(k)
17609 .cn_stride(19)
17610 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17611 }
17612 }
17613 }
17614
17615 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16_subtile) {
17616 TEST_REQUIRES_ARM_NEON;
17617 for (uint32_t n = 32; n <= 48; n += 16) {
17618 for (size_t k = 1; k <= 40; k += 9) {
17619 for (uint32_t m = 1; m <= 1; m++) {
17620 GemmMicrokernelTester()
17621 .mr(1)
17622 .nr(16)
17623 .kr(8)
17624 .sr(1)
17625 .m(m)
17626 .n(n)
17627 .k(k)
17628 .iterations(1)
17629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17630 }
17631 }
17632 }
17633 }
17634
17635 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, small_kernel) {
17636 TEST_REQUIRES_ARM_NEON;
17637 for (size_t k = 1; k <= 40; k += 9) {
17638 GemmMicrokernelTester()
17639 .mr(1)
17640 .nr(16)
17641 .kr(8)
17642 .sr(1)
17643 .m(1)
17644 .n(16)
17645 .k(k)
17646 .ks(3)
17647 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17648 }
17649 }
17650
17651 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, small_kernel_subtile) {
17652 TEST_REQUIRES_ARM_NEON;
17653 for (size_t k = 1; k <= 40; k += 9) {
17654 for (uint32_t n = 1; n <= 16; n++) {
17655 for (uint32_t m = 1; m <= 1; m++) {
17656 GemmMicrokernelTester()
17657 .mr(1)
17658 .nr(16)
17659 .kr(8)
17660 .sr(1)
17661 .m(m)
17662 .n(n)
17663 .k(k)
17664 .ks(3)
17665 .iterations(1)
17666 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17667 }
17668 }
17669 }
17670 }
17671
17672 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_gt_16_small_kernel) {
17673 TEST_REQUIRES_ARM_NEON;
17674 for (uint32_t n = 17; n < 32; n++) {
17675 for (size_t k = 1; k <= 40; k += 9) {
17676 GemmMicrokernelTester()
17677 .mr(1)
17678 .nr(16)
17679 .kr(8)
17680 .sr(1)
17681 .m(1)
17682 .n(n)
17683 .k(k)
17684 .ks(3)
17685 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17686 }
17687 }
17688 }
17689
17690 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, n_div_16_small_kernel) {
17691 TEST_REQUIRES_ARM_NEON;
17692 for (uint32_t n = 32; n <= 48; n += 16) {
17693 for (size_t k = 1; k <= 40; k += 9) {
17694 GemmMicrokernelTester()
17695 .mr(1)
17696 .nr(16)
17697 .kr(8)
17698 .sr(1)
17699 .m(1)
17700 .n(n)
17701 .k(k)
17702 .ks(3)
17703 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17704 }
17705 }
17706 }
17707
17708 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, strided_cm_subtile) {
17709 TEST_REQUIRES_ARM_NEON;
17710 for (size_t k = 1; k <= 40; k += 9) {
17711 for (uint32_t n = 1; n <= 16; n++) {
17712 for (uint32_t m = 1; m <= 1; m++) {
17713 GemmMicrokernelTester()
17714 .mr(1)
17715 .nr(16)
17716 .kr(8)
17717 .sr(1)
17718 .m(m)
17719 .n(n)
17720 .k(k)
17721 .cm_stride(19)
17722 .iterations(1)
17723 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17724 }
17725 }
17726 }
17727 }
17728
17729 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, a_offset) {
17730 TEST_REQUIRES_ARM_NEON;
17731 for (size_t k = 1; k <= 40; k += 9) {
17732 GemmMicrokernelTester()
17733 .mr(1)
17734 .nr(16)
17735 .kr(8)
17736 .sr(1)
17737 .m(1)
17738 .n(16)
17739 .k(k)
17740 .ks(3)
17741 .a_offset(43)
17742 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17743 }
17744 }
17745
17746 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, zero) {
17747 TEST_REQUIRES_ARM_NEON;
17748 for (size_t k = 1; k <= 40; k += 9) {
17749 for (uint32_t mz = 0; mz < 1; mz++) {
17750 GemmMicrokernelTester()
17751 .mr(1)
17752 .nr(16)
17753 .kr(8)
17754 .sr(1)
17755 .m(1)
17756 .n(16)
17757 .k(k)
17758 .ks(3)
17759 .a_offset(43)
17760 .zero_index(mz)
17761 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17762 }
17763 }
17764 }
17765
17766 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, qmin) {
17767 TEST_REQUIRES_ARM_NEON;
17768 GemmMicrokernelTester()
17769 .mr(1)
17770 .nr(16)
17771 .kr(8)
17772 .sr(1)
17773 .m(1)
17774 .n(16)
17775 .k(8)
17776 .qmin(128)
17777 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17778 }
17779
17780 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, qmax) {
17781 TEST_REQUIRES_ARM_NEON;
17782 GemmMicrokernelTester()
17783 .mr(1)
17784 .nr(16)
17785 .kr(8)
17786 .sr(1)
17787 .m(1)
17788 .n(16)
17789 .k(8)
17790 .qmax(128)
17791 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17792 }
17793
17794 TEST(QS8_IGEMM_MINMAX_RNDNU_1X16C8__NEON_MULL, strided_cm) {
17795 TEST_REQUIRES_ARM_NEON;
17796 GemmMicrokernelTester()
17797 .mr(1)
17798 .nr(16)
17799 .kr(8)
17800 .sr(1)
17801 .m(1)
17802 .n(16)
17803 .k(8)
17804 .cm_stride(19)
17805 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_1x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17806 }
17807#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
17808
17809
17810#if XNN_ARCH_ARM || XNN_ARCH_ARM64
17811 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_eq_8) {
17812 TEST_REQUIRES_ARM_NEON;
17813 GemmMicrokernelTester()
17814 .mr(2)
17815 .nr(16)
17816 .kr(8)
17817 .sr(1)
17818 .m(2)
17819 .n(16)
17820 .k(8)
17821 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17822 }
17823
17824 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, strided_cn) {
17825 TEST_REQUIRES_ARM_NEON;
17826 GemmMicrokernelTester()
17827 .mr(2)
17828 .nr(16)
17829 .kr(8)
17830 .sr(1)
17831 .m(2)
17832 .n(16)
17833 .k(8)
17834 .cn_stride(19)
17835 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17836 }
17837
17838 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_eq_8_subtile) {
17839 TEST_REQUIRES_ARM_NEON;
17840 for (uint32_t n = 1; n <= 16; n++) {
17841 for (uint32_t m = 1; m <= 2; m++) {
17842 GemmMicrokernelTester()
17843 .mr(2)
17844 .nr(16)
17845 .kr(8)
17846 .sr(1)
17847 .m(m)
17848 .n(n)
17849 .k(8)
17850 .iterations(1)
17851 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17852 }
17853 }
17854 }
17855
17856 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_eq_8_subtile_m) {
17857 TEST_REQUIRES_ARM_NEON;
17858 for (uint32_t m = 1; m <= 2; m++) {
17859 GemmMicrokernelTester()
17860 .mr(2)
17861 .nr(16)
17862 .kr(8)
17863 .sr(1)
17864 .m(m)
17865 .n(16)
17866 .k(8)
17867 .iterations(1)
17868 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17869 }
17870 }
17871
17872 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_eq_8_subtile_n) {
17873 TEST_REQUIRES_ARM_NEON;
17874 for (uint32_t n = 1; n <= 16; n++) {
17875 GemmMicrokernelTester()
17876 .mr(2)
17877 .nr(16)
17878 .kr(8)
17879 .sr(1)
17880 .m(2)
17881 .n(n)
17882 .k(8)
17883 .iterations(1)
17884 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17885 }
17886 }
17887
17888 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_lt_8) {
17889 TEST_REQUIRES_ARM_NEON;
17890 for (size_t k = 1; k < 8; k++) {
17891 GemmMicrokernelTester()
17892 .mr(2)
17893 .nr(16)
17894 .kr(8)
17895 .sr(1)
17896 .m(2)
17897 .n(16)
17898 .k(k)
17899 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17900 }
17901 }
17902
17903 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_lt_8_subtile) {
17904 TEST_REQUIRES_ARM_NEON;
17905 for (size_t k = 1; k < 8; k++) {
17906 for (uint32_t n = 1; n <= 16; n++) {
17907 for (uint32_t m = 1; m <= 2; m++) {
17908 GemmMicrokernelTester()
17909 .mr(2)
17910 .nr(16)
17911 .kr(8)
17912 .sr(1)
17913 .m(m)
17914 .n(n)
17915 .k(k)
17916 .iterations(1)
17917 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17918 }
17919 }
17920 }
17921 }
17922
17923 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_gt_8) {
17924 TEST_REQUIRES_ARM_NEON;
17925 for (size_t k = 9; k < 16; k++) {
17926 GemmMicrokernelTester()
17927 .mr(2)
17928 .nr(16)
17929 .kr(8)
17930 .sr(1)
17931 .m(2)
17932 .n(16)
17933 .k(k)
17934 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17935 }
17936 }
17937
17938 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_gt_8_subtile) {
17939 TEST_REQUIRES_ARM_NEON;
17940 for (size_t k = 9; k < 16; k++) {
17941 for (uint32_t n = 1; n <= 16; n++) {
17942 for (uint32_t m = 1; m <= 2; m++) {
17943 GemmMicrokernelTester()
17944 .mr(2)
17945 .nr(16)
17946 .kr(8)
17947 .sr(1)
17948 .m(m)
17949 .n(n)
17950 .k(k)
17951 .iterations(1)
17952 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17953 }
17954 }
17955 }
17956 }
17957
17958 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_div_8) {
17959 TEST_REQUIRES_ARM_NEON;
17960 for (size_t k = 16; k <= 80; k += 8) {
17961 GemmMicrokernelTester()
17962 .mr(2)
17963 .nr(16)
17964 .kr(8)
17965 .sr(1)
17966 .m(2)
17967 .n(16)
17968 .k(k)
17969 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17970 }
17971 }
17972
17973 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, k_div_8_subtile) {
17974 TEST_REQUIRES_ARM_NEON;
17975 for (size_t k = 16; k <= 80; k += 8) {
17976 for (uint32_t n = 1; n <= 16; n++) {
17977 for (uint32_t m = 1; m <= 2; m++) {
17978 GemmMicrokernelTester()
17979 .mr(2)
17980 .nr(16)
17981 .kr(8)
17982 .sr(1)
17983 .m(m)
17984 .n(n)
17985 .k(k)
17986 .iterations(1)
17987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
17988 }
17989 }
17990 }
17991 }
17992
17993 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_gt_16) {
17994 TEST_REQUIRES_ARM_NEON;
17995 for (uint32_t n = 17; n < 32; n++) {
17996 for (size_t k = 1; k <= 40; k += 9) {
17997 GemmMicrokernelTester()
17998 .mr(2)
17999 .nr(16)
18000 .kr(8)
18001 .sr(1)
18002 .m(2)
18003 .n(n)
18004 .k(k)
18005 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18006 }
18007 }
18008 }
18009
18010 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_gt_16_strided_cn) {
18011 TEST_REQUIRES_ARM_NEON;
18012 for (uint32_t n = 17; n < 32; n++) {
18013 for (size_t k = 1; k <= 40; k += 9) {
18014 GemmMicrokernelTester()
18015 .mr(2)
18016 .nr(16)
18017 .kr(8)
18018 .sr(1)
18019 .m(2)
18020 .n(n)
18021 .k(k)
18022 .cn_stride(19)
18023 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18024 }
18025 }
18026 }
18027
18028 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_gt_16_subtile) {
18029 TEST_REQUIRES_ARM_NEON;
18030 for (uint32_t n = 17; n < 32; n++) {
18031 for (size_t k = 1; k <= 40; k += 9) {
18032 for (uint32_t m = 1; m <= 2; m++) {
18033 GemmMicrokernelTester()
18034 .mr(2)
18035 .nr(16)
18036 .kr(8)
18037 .sr(1)
18038 .m(m)
18039 .n(n)
18040 .k(k)
18041 .iterations(1)
18042 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18043 }
18044 }
18045 }
18046 }
18047
18048 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_div_16) {
18049 TEST_REQUIRES_ARM_NEON;
18050 for (uint32_t n = 32; n <= 48; n += 16) {
18051 for (size_t k = 1; k <= 40; k += 9) {
18052 GemmMicrokernelTester()
18053 .mr(2)
18054 .nr(16)
18055 .kr(8)
18056 .sr(1)
18057 .m(2)
18058 .n(n)
18059 .k(k)
18060 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18061 }
18062 }
18063 }
18064
18065 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_div_16_strided_cn) {
18066 TEST_REQUIRES_ARM_NEON;
18067 for (uint32_t n = 32; n <= 48; n += 16) {
18068 for (size_t k = 1; k <= 40; k += 9) {
18069 GemmMicrokernelTester()
18070 .mr(2)
18071 .nr(16)
18072 .kr(8)
18073 .sr(1)
18074 .m(2)
18075 .n(n)
18076 .k(k)
18077 .cn_stride(19)
18078 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18079 }
18080 }
18081 }
18082
18083 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_div_16_subtile) {
18084 TEST_REQUIRES_ARM_NEON;
18085 for (uint32_t n = 32; n <= 48; n += 16) {
18086 for (size_t k = 1; k <= 40; k += 9) {
18087 for (uint32_t m = 1; m <= 2; m++) {
18088 GemmMicrokernelTester()
18089 .mr(2)
18090 .nr(16)
18091 .kr(8)
18092 .sr(1)
18093 .m(m)
18094 .n(n)
18095 .k(k)
18096 .iterations(1)
18097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18098 }
18099 }
18100 }
18101 }
18102
18103 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, small_kernel) {
18104 TEST_REQUIRES_ARM_NEON;
18105 for (size_t k = 1; k <= 40; k += 9) {
18106 GemmMicrokernelTester()
18107 .mr(2)
18108 .nr(16)
18109 .kr(8)
18110 .sr(1)
18111 .m(2)
18112 .n(16)
18113 .k(k)
18114 .ks(3)
18115 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18116 }
18117 }
18118
18119 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, small_kernel_subtile) {
18120 TEST_REQUIRES_ARM_NEON;
18121 for (size_t k = 1; k <= 40; k += 9) {
18122 for (uint32_t n = 1; n <= 16; n++) {
18123 for (uint32_t m = 1; m <= 2; m++) {
18124 GemmMicrokernelTester()
18125 .mr(2)
18126 .nr(16)
18127 .kr(8)
18128 .sr(1)
18129 .m(m)
18130 .n(n)
18131 .k(k)
18132 .ks(3)
18133 .iterations(1)
18134 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18135 }
18136 }
18137 }
18138 }
18139
18140 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_gt_16_small_kernel) {
18141 TEST_REQUIRES_ARM_NEON;
18142 for (uint32_t n = 17; n < 32; n++) {
18143 for (size_t k = 1; k <= 40; k += 9) {
18144 GemmMicrokernelTester()
18145 .mr(2)
18146 .nr(16)
18147 .kr(8)
18148 .sr(1)
18149 .m(2)
18150 .n(n)
18151 .k(k)
18152 .ks(3)
18153 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18154 }
18155 }
18156 }
18157
18158 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, n_div_16_small_kernel) {
18159 TEST_REQUIRES_ARM_NEON;
18160 for (uint32_t n = 32; n <= 48; n += 16) {
18161 for (size_t k = 1; k <= 40; k += 9) {
18162 GemmMicrokernelTester()
18163 .mr(2)
18164 .nr(16)
18165 .kr(8)
18166 .sr(1)
18167 .m(2)
18168 .n(n)
18169 .k(k)
18170 .ks(3)
18171 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18172 }
18173 }
18174 }
18175
18176 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, strided_cm_subtile) {
18177 TEST_REQUIRES_ARM_NEON;
18178 for (size_t k = 1; k <= 40; k += 9) {
18179 for (uint32_t n = 1; n <= 16; n++) {
18180 for (uint32_t m = 1; m <= 2; m++) {
18181 GemmMicrokernelTester()
18182 .mr(2)
18183 .nr(16)
18184 .kr(8)
18185 .sr(1)
18186 .m(m)
18187 .n(n)
18188 .k(k)
18189 .cm_stride(19)
18190 .iterations(1)
18191 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18192 }
18193 }
18194 }
18195 }
18196
18197 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, a_offset) {
18198 TEST_REQUIRES_ARM_NEON;
18199 for (size_t k = 1; k <= 40; k += 9) {
18200 GemmMicrokernelTester()
18201 .mr(2)
18202 .nr(16)
18203 .kr(8)
18204 .sr(1)
18205 .m(2)
18206 .n(16)
18207 .k(k)
18208 .ks(3)
18209 .a_offset(83)
18210 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18211 }
18212 }
18213
18214 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, zero) {
18215 TEST_REQUIRES_ARM_NEON;
18216 for (size_t k = 1; k <= 40; k += 9) {
18217 for (uint32_t mz = 0; mz < 2; mz++) {
18218 GemmMicrokernelTester()
18219 .mr(2)
18220 .nr(16)
18221 .kr(8)
18222 .sr(1)
18223 .m(2)
18224 .n(16)
18225 .k(k)
18226 .ks(3)
18227 .a_offset(83)
18228 .zero_index(mz)
18229 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18230 }
18231 }
18232 }
18233
18234 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, qmin) {
18235 TEST_REQUIRES_ARM_NEON;
18236 GemmMicrokernelTester()
18237 .mr(2)
18238 .nr(16)
18239 .kr(8)
18240 .sr(1)
18241 .m(2)
18242 .n(16)
18243 .k(8)
18244 .qmin(128)
18245 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18246 }
18247
18248 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, qmax) {
18249 TEST_REQUIRES_ARM_NEON;
18250 GemmMicrokernelTester()
18251 .mr(2)
18252 .nr(16)
18253 .kr(8)
18254 .sr(1)
18255 .m(2)
18256 .n(16)
18257 .k(8)
18258 .qmax(128)
18259 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18260 }
18261
18262 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16C8__NEON_MULL, strided_cm) {
18263 TEST_REQUIRES_ARM_NEON;
18264 GemmMicrokernelTester()
18265 .mr(2)
18266 .nr(16)
18267 .kr(8)
18268 .sr(1)
18269 .m(2)
18270 .n(16)
18271 .k(8)
18272 .cm_stride(19)
18273 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18274 }
18275#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
18276
18277
18278#if XNN_ARCH_ARM || XNN_ARCH_ARM64
18279 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_eq_8) {
18280 TEST_REQUIRES_ARM_NEON;
18281 GemmMicrokernelTester()
18282 .mr(4)
18283 .nr(16)
18284 .kr(8)
18285 .sr(1)
18286 .m(4)
18287 .n(16)
18288 .k(8)
18289 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18290 }
18291
18292 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, strided_cn) {
18293 TEST_REQUIRES_ARM_NEON;
18294 GemmMicrokernelTester()
18295 .mr(4)
18296 .nr(16)
18297 .kr(8)
18298 .sr(1)
18299 .m(4)
18300 .n(16)
18301 .k(8)
18302 .cn_stride(19)
18303 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18304 }
18305
18306 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_eq_8_subtile) {
18307 TEST_REQUIRES_ARM_NEON;
18308 for (uint32_t n = 1; n <= 16; n++) {
18309 for (uint32_t m = 1; m <= 4; m++) {
18310 GemmMicrokernelTester()
18311 .mr(4)
18312 .nr(16)
18313 .kr(8)
18314 .sr(1)
18315 .m(m)
18316 .n(n)
18317 .k(8)
18318 .iterations(1)
18319 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18320 }
18321 }
18322 }
18323
18324 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_eq_8_subtile_m) {
18325 TEST_REQUIRES_ARM_NEON;
18326 for (uint32_t m = 1; m <= 4; m++) {
18327 GemmMicrokernelTester()
18328 .mr(4)
18329 .nr(16)
18330 .kr(8)
18331 .sr(1)
18332 .m(m)
18333 .n(16)
18334 .k(8)
18335 .iterations(1)
18336 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18337 }
18338 }
18339
18340 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_eq_8_subtile_n) {
18341 TEST_REQUIRES_ARM_NEON;
18342 for (uint32_t n = 1; n <= 16; n++) {
18343 GemmMicrokernelTester()
18344 .mr(4)
18345 .nr(16)
18346 .kr(8)
18347 .sr(1)
18348 .m(4)
18349 .n(n)
18350 .k(8)
18351 .iterations(1)
18352 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18353 }
18354 }
18355
18356 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_lt_8) {
18357 TEST_REQUIRES_ARM_NEON;
18358 for (size_t k = 1; k < 8; k++) {
18359 GemmMicrokernelTester()
18360 .mr(4)
18361 .nr(16)
18362 .kr(8)
18363 .sr(1)
18364 .m(4)
18365 .n(16)
18366 .k(k)
18367 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18368 }
18369 }
18370
18371 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_lt_8_subtile) {
18372 TEST_REQUIRES_ARM_NEON;
18373 for (size_t k = 1; k < 8; k++) {
18374 for (uint32_t n = 1; n <= 16; n++) {
18375 for (uint32_t m = 1; m <= 4; m++) {
18376 GemmMicrokernelTester()
18377 .mr(4)
18378 .nr(16)
18379 .kr(8)
18380 .sr(1)
18381 .m(m)
18382 .n(n)
18383 .k(k)
18384 .iterations(1)
18385 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18386 }
18387 }
18388 }
18389 }
18390
18391 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_gt_8) {
18392 TEST_REQUIRES_ARM_NEON;
18393 for (size_t k = 9; k < 16; k++) {
18394 GemmMicrokernelTester()
18395 .mr(4)
18396 .nr(16)
18397 .kr(8)
18398 .sr(1)
18399 .m(4)
18400 .n(16)
18401 .k(k)
18402 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18403 }
18404 }
18405
18406 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_gt_8_subtile) {
18407 TEST_REQUIRES_ARM_NEON;
18408 for (size_t k = 9; k < 16; k++) {
18409 for (uint32_t n = 1; n <= 16; n++) {
18410 for (uint32_t m = 1; m <= 4; m++) {
18411 GemmMicrokernelTester()
18412 .mr(4)
18413 .nr(16)
18414 .kr(8)
18415 .sr(1)
18416 .m(m)
18417 .n(n)
18418 .k(k)
18419 .iterations(1)
18420 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18421 }
18422 }
18423 }
18424 }
18425
18426 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_div_8) {
18427 TEST_REQUIRES_ARM_NEON;
18428 for (size_t k = 16; k <= 80; k += 8) {
18429 GemmMicrokernelTester()
18430 .mr(4)
18431 .nr(16)
18432 .kr(8)
18433 .sr(1)
18434 .m(4)
18435 .n(16)
18436 .k(k)
18437 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18438 }
18439 }
18440
18441 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, k_div_8_subtile) {
18442 TEST_REQUIRES_ARM_NEON;
18443 for (size_t k = 16; k <= 80; k += 8) {
18444 for (uint32_t n = 1; n <= 16; n++) {
18445 for (uint32_t m = 1; m <= 4; m++) {
18446 GemmMicrokernelTester()
18447 .mr(4)
18448 .nr(16)
18449 .kr(8)
18450 .sr(1)
18451 .m(m)
18452 .n(n)
18453 .k(k)
18454 .iterations(1)
18455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18456 }
18457 }
18458 }
18459 }
18460
18461 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_gt_16) {
18462 TEST_REQUIRES_ARM_NEON;
18463 for (uint32_t n = 17; n < 32; n++) {
18464 for (size_t k = 1; k <= 40; k += 9) {
18465 GemmMicrokernelTester()
18466 .mr(4)
18467 .nr(16)
18468 .kr(8)
18469 .sr(1)
18470 .m(4)
18471 .n(n)
18472 .k(k)
18473 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18474 }
18475 }
18476 }
18477
18478 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_gt_16_strided_cn) {
18479 TEST_REQUIRES_ARM_NEON;
18480 for (uint32_t n = 17; n < 32; n++) {
18481 for (size_t k = 1; k <= 40; k += 9) {
18482 GemmMicrokernelTester()
18483 .mr(4)
18484 .nr(16)
18485 .kr(8)
18486 .sr(1)
18487 .m(4)
18488 .n(n)
18489 .k(k)
18490 .cn_stride(19)
18491 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18492 }
18493 }
18494 }
18495
18496 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_gt_16_subtile) {
18497 TEST_REQUIRES_ARM_NEON;
18498 for (uint32_t n = 17; n < 32; n++) {
18499 for (size_t k = 1; k <= 40; k += 9) {
18500 for (uint32_t m = 1; m <= 4; m++) {
18501 GemmMicrokernelTester()
18502 .mr(4)
18503 .nr(16)
18504 .kr(8)
18505 .sr(1)
18506 .m(m)
18507 .n(n)
18508 .k(k)
18509 .iterations(1)
18510 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18511 }
18512 }
18513 }
18514 }
18515
18516 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_div_16) {
18517 TEST_REQUIRES_ARM_NEON;
18518 for (uint32_t n = 32; n <= 48; n += 16) {
18519 for (size_t k = 1; k <= 40; k += 9) {
18520 GemmMicrokernelTester()
18521 .mr(4)
18522 .nr(16)
18523 .kr(8)
18524 .sr(1)
18525 .m(4)
18526 .n(n)
18527 .k(k)
18528 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18529 }
18530 }
18531 }
18532
18533 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_div_16_strided_cn) {
18534 TEST_REQUIRES_ARM_NEON;
18535 for (uint32_t n = 32; n <= 48; n += 16) {
18536 for (size_t k = 1; k <= 40; k += 9) {
18537 GemmMicrokernelTester()
18538 .mr(4)
18539 .nr(16)
18540 .kr(8)
18541 .sr(1)
18542 .m(4)
18543 .n(n)
18544 .k(k)
18545 .cn_stride(19)
18546 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18547 }
18548 }
18549 }
18550
18551 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_div_16_subtile) {
18552 TEST_REQUIRES_ARM_NEON;
18553 for (uint32_t n = 32; n <= 48; n += 16) {
18554 for (size_t k = 1; k <= 40; k += 9) {
18555 for (uint32_t m = 1; m <= 4; m++) {
18556 GemmMicrokernelTester()
18557 .mr(4)
18558 .nr(16)
18559 .kr(8)
18560 .sr(1)
18561 .m(m)
18562 .n(n)
18563 .k(k)
18564 .iterations(1)
18565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18566 }
18567 }
18568 }
18569 }
18570
18571 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, small_kernel) {
18572 TEST_REQUIRES_ARM_NEON;
18573 for (size_t k = 1; k <= 40; k += 9) {
18574 GemmMicrokernelTester()
18575 .mr(4)
18576 .nr(16)
18577 .kr(8)
18578 .sr(1)
18579 .m(4)
18580 .n(16)
18581 .k(k)
18582 .ks(3)
18583 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18584 }
18585 }
18586
18587 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, small_kernel_subtile) {
18588 TEST_REQUIRES_ARM_NEON;
18589 for (size_t k = 1; k <= 40; k += 9) {
18590 for (uint32_t n = 1; n <= 16; n++) {
18591 for (uint32_t m = 1; m <= 4; m++) {
18592 GemmMicrokernelTester()
18593 .mr(4)
18594 .nr(16)
18595 .kr(8)
18596 .sr(1)
18597 .m(m)
18598 .n(n)
18599 .k(k)
18600 .ks(3)
18601 .iterations(1)
18602 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18603 }
18604 }
18605 }
18606 }
18607
18608 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_gt_16_small_kernel) {
18609 TEST_REQUIRES_ARM_NEON;
18610 for (uint32_t n = 17; n < 32; n++) {
18611 for (size_t k = 1; k <= 40; k += 9) {
18612 GemmMicrokernelTester()
18613 .mr(4)
18614 .nr(16)
18615 .kr(8)
18616 .sr(1)
18617 .m(4)
18618 .n(n)
18619 .k(k)
18620 .ks(3)
18621 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18622 }
18623 }
18624 }
18625
18626 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, n_div_16_small_kernel) {
18627 TEST_REQUIRES_ARM_NEON;
18628 for (uint32_t n = 32; n <= 48; n += 16) {
18629 for (size_t k = 1; k <= 40; k += 9) {
18630 GemmMicrokernelTester()
18631 .mr(4)
18632 .nr(16)
18633 .kr(8)
18634 .sr(1)
18635 .m(4)
18636 .n(n)
18637 .k(k)
18638 .ks(3)
18639 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18640 }
18641 }
18642 }
18643
18644 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, strided_cm_subtile) {
18645 TEST_REQUIRES_ARM_NEON;
18646 for (size_t k = 1; k <= 40; k += 9) {
18647 for (uint32_t n = 1; n <= 16; n++) {
18648 for (uint32_t m = 1; m <= 4; m++) {
18649 GemmMicrokernelTester()
18650 .mr(4)
18651 .nr(16)
18652 .kr(8)
18653 .sr(1)
18654 .m(m)
18655 .n(n)
18656 .k(k)
18657 .cm_stride(19)
18658 .iterations(1)
18659 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18660 }
18661 }
18662 }
18663 }
18664
18665 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, a_offset) {
18666 TEST_REQUIRES_ARM_NEON;
18667 for (size_t k = 1; k <= 40; k += 9) {
18668 GemmMicrokernelTester()
18669 .mr(4)
18670 .nr(16)
18671 .kr(8)
18672 .sr(1)
18673 .m(4)
18674 .n(16)
18675 .k(k)
18676 .ks(3)
18677 .a_offset(163)
18678 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18679 }
18680 }
18681
18682 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, zero) {
18683 TEST_REQUIRES_ARM_NEON;
18684 for (size_t k = 1; k <= 40; k += 9) {
18685 for (uint32_t mz = 0; mz < 4; mz++) {
18686 GemmMicrokernelTester()
18687 .mr(4)
18688 .nr(16)
18689 .kr(8)
18690 .sr(1)
18691 .m(4)
18692 .n(16)
18693 .k(k)
18694 .ks(3)
18695 .a_offset(163)
18696 .zero_index(mz)
18697 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18698 }
18699 }
18700 }
18701
18702 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, qmin) {
18703 TEST_REQUIRES_ARM_NEON;
18704 GemmMicrokernelTester()
18705 .mr(4)
18706 .nr(16)
18707 .kr(8)
18708 .sr(1)
18709 .m(4)
18710 .n(16)
18711 .k(8)
18712 .qmin(128)
18713 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18714 }
18715
18716 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, qmax) {
18717 TEST_REQUIRES_ARM_NEON;
18718 GemmMicrokernelTester()
18719 .mr(4)
18720 .nr(16)
18721 .kr(8)
18722 .sr(1)
18723 .m(4)
18724 .n(16)
18725 .k(8)
18726 .qmax(128)
18727 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18728 }
18729
18730 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C8__NEON_MULL, strided_cm) {
18731 TEST_REQUIRES_ARM_NEON;
18732 GemmMicrokernelTester()
18733 .mr(4)
18734 .nr(16)
18735 .kr(8)
18736 .sr(1)
18737 .m(4)
18738 .n(16)
18739 .k(8)
18740 .cm_stride(19)
18741 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c8__neon_mull, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18742 }
18743#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
18744
18745
18746#if XNN_ARCH_ARM || XNN_ARCH_ARM64
18747 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_eq_16) {
18748 TEST_REQUIRES_ARM_NEON;
18749 GemmMicrokernelTester()
18750 .mr(3)
18751 .nr(16)
18752 .kr(16)
18753 .sr(1)
18754 .m(3)
18755 .n(16)
18756 .k(16)
18757 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18758 }
18759
18760 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, strided_cn) {
18761 TEST_REQUIRES_ARM_NEON;
18762 GemmMicrokernelTester()
18763 .mr(3)
18764 .nr(16)
18765 .kr(16)
18766 .sr(1)
18767 .m(3)
18768 .n(16)
18769 .k(16)
18770 .cn_stride(19)
18771 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18772 }
18773
18774 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_eq_16_subtile) {
18775 TEST_REQUIRES_ARM_NEON;
18776 for (uint32_t n = 1; n <= 16; n++) {
18777 for (uint32_t m = 1; m <= 3; m++) {
18778 GemmMicrokernelTester()
18779 .mr(3)
18780 .nr(16)
18781 .kr(16)
18782 .sr(1)
18783 .m(m)
18784 .n(n)
18785 .k(16)
18786 .iterations(1)
18787 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18788 }
18789 }
18790 }
18791
18792 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_eq_16_subtile_m) {
18793 TEST_REQUIRES_ARM_NEON;
18794 for (uint32_t m = 1; m <= 3; m++) {
18795 GemmMicrokernelTester()
18796 .mr(3)
18797 .nr(16)
18798 .kr(16)
18799 .sr(1)
18800 .m(m)
18801 .n(16)
18802 .k(16)
18803 .iterations(1)
18804 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18805 }
18806 }
18807
18808 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_eq_16_subtile_n) {
18809 TEST_REQUIRES_ARM_NEON;
18810 for (uint32_t n = 1; n <= 16; n++) {
18811 GemmMicrokernelTester()
18812 .mr(3)
18813 .nr(16)
18814 .kr(16)
18815 .sr(1)
18816 .m(3)
18817 .n(n)
18818 .k(16)
18819 .iterations(1)
18820 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18821 }
18822 }
18823
18824 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_lt_16) {
18825 TEST_REQUIRES_ARM_NEON;
18826 for (size_t k = 1; k < 16; k++) {
18827 GemmMicrokernelTester()
18828 .mr(3)
18829 .nr(16)
18830 .kr(16)
18831 .sr(1)
18832 .m(3)
18833 .n(16)
18834 .k(k)
18835 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18836 }
18837 }
18838
18839 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_lt_16_subtile) {
18840 TEST_REQUIRES_ARM_NEON;
18841 for (size_t k = 1; k < 16; k++) {
18842 for (uint32_t n = 1; n <= 16; n++) {
18843 for (uint32_t m = 1; m <= 3; m++) {
18844 GemmMicrokernelTester()
18845 .mr(3)
18846 .nr(16)
18847 .kr(16)
18848 .sr(1)
18849 .m(m)
18850 .n(n)
18851 .k(k)
18852 .iterations(1)
18853 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18854 }
18855 }
18856 }
18857 }
18858
18859 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_gt_16) {
18860 TEST_REQUIRES_ARM_NEON;
18861 for (size_t k = 17; k < 32; k++) {
18862 GemmMicrokernelTester()
18863 .mr(3)
18864 .nr(16)
18865 .kr(16)
18866 .sr(1)
18867 .m(3)
18868 .n(16)
18869 .k(k)
18870 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18871 }
18872 }
18873
18874 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_gt_16_subtile) {
18875 TEST_REQUIRES_ARM_NEON;
18876 for (size_t k = 17; k < 32; k++) {
18877 for (uint32_t n = 1; n <= 16; n++) {
18878 for (uint32_t m = 1; m <= 3; m++) {
18879 GemmMicrokernelTester()
18880 .mr(3)
18881 .nr(16)
18882 .kr(16)
18883 .sr(1)
18884 .m(m)
18885 .n(n)
18886 .k(k)
18887 .iterations(1)
18888 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18889 }
18890 }
18891 }
18892 }
18893
18894 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_div_16) {
18895 TEST_REQUIRES_ARM_NEON;
18896 for (size_t k = 32; k <= 160; k += 16) {
18897 GemmMicrokernelTester()
18898 .mr(3)
18899 .nr(16)
18900 .kr(16)
18901 .sr(1)
18902 .m(3)
18903 .n(16)
18904 .k(k)
18905 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18906 }
18907 }
18908
18909 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, k_div_16_subtile) {
18910 TEST_REQUIRES_ARM_NEON;
18911 for (size_t k = 32; k <= 160; k += 16) {
18912 for (uint32_t n = 1; n <= 16; n++) {
18913 for (uint32_t m = 1; m <= 3; m++) {
18914 GemmMicrokernelTester()
18915 .mr(3)
18916 .nr(16)
18917 .kr(16)
18918 .sr(1)
18919 .m(m)
18920 .n(n)
18921 .k(k)
18922 .iterations(1)
18923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18924 }
18925 }
18926 }
18927 }
18928
18929 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_gt_16) {
18930 TEST_REQUIRES_ARM_NEON;
18931 for (uint32_t n = 17; n < 32; n++) {
18932 for (size_t k = 1; k <= 80; k += 17) {
18933 GemmMicrokernelTester()
18934 .mr(3)
18935 .nr(16)
18936 .kr(16)
18937 .sr(1)
18938 .m(3)
18939 .n(n)
18940 .k(k)
18941 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18942 }
18943 }
18944 }
18945
18946 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_gt_16_strided_cn) {
18947 TEST_REQUIRES_ARM_NEON;
18948 for (uint32_t n = 17; n < 32; n++) {
18949 for (size_t k = 1; k <= 80; k += 17) {
18950 GemmMicrokernelTester()
18951 .mr(3)
18952 .nr(16)
18953 .kr(16)
18954 .sr(1)
18955 .m(3)
18956 .n(n)
18957 .k(k)
18958 .cn_stride(19)
18959 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18960 }
18961 }
18962 }
18963
18964 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_gt_16_subtile) {
18965 TEST_REQUIRES_ARM_NEON;
18966 for (uint32_t n = 17; n < 32; n++) {
18967 for (size_t k = 1; k <= 80; k += 17) {
18968 for (uint32_t m = 1; m <= 3; m++) {
18969 GemmMicrokernelTester()
18970 .mr(3)
18971 .nr(16)
18972 .kr(16)
18973 .sr(1)
18974 .m(m)
18975 .n(n)
18976 .k(k)
18977 .iterations(1)
18978 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18979 }
18980 }
18981 }
18982 }
18983
18984 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_div_16) {
18985 TEST_REQUIRES_ARM_NEON;
18986 for (uint32_t n = 32; n <= 48; n += 16) {
18987 for (size_t k = 1; k <= 80; k += 17) {
18988 GemmMicrokernelTester()
18989 .mr(3)
18990 .nr(16)
18991 .kr(16)
18992 .sr(1)
18993 .m(3)
18994 .n(n)
18995 .k(k)
18996 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
18997 }
18998 }
18999 }
19000
19001 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_div_16_strided_cn) {
19002 TEST_REQUIRES_ARM_NEON;
19003 for (uint32_t n = 32; n <= 48; n += 16) {
19004 for (size_t k = 1; k <= 80; k += 17) {
19005 GemmMicrokernelTester()
19006 .mr(3)
19007 .nr(16)
19008 .kr(16)
19009 .sr(1)
19010 .m(3)
19011 .n(n)
19012 .k(k)
19013 .cn_stride(19)
19014 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19015 }
19016 }
19017 }
19018
19019 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_div_16_subtile) {
19020 TEST_REQUIRES_ARM_NEON;
19021 for (uint32_t n = 32; n <= 48; n += 16) {
19022 for (size_t k = 1; k <= 80; k += 17) {
19023 for (uint32_t m = 1; m <= 3; m++) {
19024 GemmMicrokernelTester()
19025 .mr(3)
19026 .nr(16)
19027 .kr(16)
19028 .sr(1)
19029 .m(m)
19030 .n(n)
19031 .k(k)
19032 .iterations(1)
19033 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19034 }
19035 }
19036 }
19037 }
19038
19039 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, small_kernel) {
19040 TEST_REQUIRES_ARM_NEON;
19041 for (size_t k = 1; k <= 80; k += 17) {
19042 GemmMicrokernelTester()
19043 .mr(3)
19044 .nr(16)
19045 .kr(16)
19046 .sr(1)
19047 .m(3)
19048 .n(16)
19049 .k(k)
19050 .ks(3)
19051 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19052 }
19053 }
19054
19055 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, small_kernel_subtile) {
19056 TEST_REQUIRES_ARM_NEON;
19057 for (size_t k = 1; k <= 80; k += 17) {
19058 for (uint32_t n = 1; n <= 16; n++) {
19059 for (uint32_t m = 1; m <= 3; m++) {
19060 GemmMicrokernelTester()
19061 .mr(3)
19062 .nr(16)
19063 .kr(16)
19064 .sr(1)
19065 .m(m)
19066 .n(n)
19067 .k(k)
19068 .ks(3)
19069 .iterations(1)
19070 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19071 }
19072 }
19073 }
19074 }
19075
19076 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_gt_16_small_kernel) {
19077 TEST_REQUIRES_ARM_NEON;
19078 for (uint32_t n = 17; n < 32; n++) {
19079 for (size_t k = 1; k <= 80; k += 17) {
19080 GemmMicrokernelTester()
19081 .mr(3)
19082 .nr(16)
19083 .kr(16)
19084 .sr(1)
19085 .m(3)
19086 .n(n)
19087 .k(k)
19088 .ks(3)
19089 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19090 }
19091 }
19092 }
19093
19094 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, n_div_16_small_kernel) {
19095 TEST_REQUIRES_ARM_NEON;
19096 for (uint32_t n = 32; n <= 48; n += 16) {
19097 for (size_t k = 1; k <= 80; k += 17) {
19098 GemmMicrokernelTester()
19099 .mr(3)
19100 .nr(16)
19101 .kr(16)
19102 .sr(1)
19103 .m(3)
19104 .n(n)
19105 .k(k)
19106 .ks(3)
19107 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19108 }
19109 }
19110 }
19111
19112 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, strided_cm_subtile) {
19113 TEST_REQUIRES_ARM_NEON;
19114 for (size_t k = 1; k <= 80; k += 17) {
19115 for (uint32_t n = 1; n <= 16; n++) {
19116 for (uint32_t m = 1; m <= 3; m++) {
19117 GemmMicrokernelTester()
19118 .mr(3)
19119 .nr(16)
19120 .kr(16)
19121 .sr(1)
19122 .m(m)
19123 .n(n)
19124 .k(k)
19125 .cm_stride(19)
19126 .iterations(1)
19127 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19128 }
19129 }
19130 }
19131 }
19132
19133 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, a_offset) {
19134 TEST_REQUIRES_ARM_NEON;
19135 for (size_t k = 1; k <= 80; k += 17) {
19136 GemmMicrokernelTester()
19137 .mr(3)
19138 .nr(16)
19139 .kr(16)
19140 .sr(1)
19141 .m(3)
19142 .n(16)
19143 .k(k)
19144 .ks(3)
19145 .a_offset(251)
19146 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19147 }
19148 }
19149
19150 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, zero) {
19151 TEST_REQUIRES_ARM_NEON;
19152 for (size_t k = 1; k <= 80; k += 17) {
19153 for (uint32_t mz = 0; mz < 3; mz++) {
19154 GemmMicrokernelTester()
19155 .mr(3)
19156 .nr(16)
19157 .kr(16)
19158 .sr(1)
19159 .m(3)
19160 .n(16)
19161 .k(k)
19162 .ks(3)
19163 .a_offset(251)
19164 .zero_index(mz)
19165 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19166 }
19167 }
19168 }
19169
19170 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, qmin) {
19171 TEST_REQUIRES_ARM_NEON;
19172 GemmMicrokernelTester()
19173 .mr(3)
19174 .nr(16)
19175 .kr(16)
19176 .sr(1)
19177 .m(3)
19178 .n(16)
19179 .k(16)
19180 .qmin(128)
19181 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19182 }
19183
19184 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, qmax) {
19185 TEST_REQUIRES_ARM_NEON;
19186 GemmMicrokernelTester()
19187 .mr(3)
19188 .nr(16)
19189 .kr(16)
19190 .sr(1)
19191 .m(3)
19192 .n(16)
19193 .k(16)
19194 .qmax(128)
19195 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19196 }
19197
19198 TEST(QS8_IGEMM_MINMAX_RNDNU_3X16C16__NEON_MLAL, strided_cm) {
19199 TEST_REQUIRES_ARM_NEON;
19200 GemmMicrokernelTester()
19201 .mr(3)
19202 .nr(16)
19203 .kr(16)
19204 .sr(1)
19205 .m(3)
19206 .n(16)
19207 .k(16)
19208 .cm_stride(19)
19209 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_3x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19210 }
19211#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
19212
19213
19214#if XNN_ARCH_ARM || XNN_ARCH_ARM64
19215 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16) {
19216 TEST_REQUIRES_ARM_NEON;
19217 GemmMicrokernelTester()
19218 .mr(4)
19219 .nr(16)
19220 .kr(16)
19221 .sr(1)
19222 .m(4)
19223 .n(16)
19224 .k(16)
19225 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19226 }
19227
19228 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, strided_cn) {
19229 TEST_REQUIRES_ARM_NEON;
19230 GemmMicrokernelTester()
19231 .mr(4)
19232 .nr(16)
19233 .kr(16)
19234 .sr(1)
19235 .m(4)
19236 .n(16)
19237 .k(16)
19238 .cn_stride(19)
19239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19240 }
19241
19242 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16_subtile) {
19243 TEST_REQUIRES_ARM_NEON;
19244 for (uint32_t n = 1; n <= 16; n++) {
19245 for (uint32_t m = 1; m <= 4; m++) {
19246 GemmMicrokernelTester()
19247 .mr(4)
19248 .nr(16)
19249 .kr(16)
19250 .sr(1)
19251 .m(m)
19252 .n(n)
19253 .k(16)
19254 .iterations(1)
19255 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19256 }
19257 }
19258 }
19259
19260 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16_subtile_m) {
19261 TEST_REQUIRES_ARM_NEON;
19262 for (uint32_t m = 1; m <= 4; m++) {
19263 GemmMicrokernelTester()
19264 .mr(4)
19265 .nr(16)
19266 .kr(16)
19267 .sr(1)
19268 .m(m)
19269 .n(16)
19270 .k(16)
19271 .iterations(1)
19272 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19273 }
19274 }
19275
19276 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_eq_16_subtile_n) {
19277 TEST_REQUIRES_ARM_NEON;
19278 for (uint32_t n = 1; n <= 16; n++) {
19279 GemmMicrokernelTester()
19280 .mr(4)
19281 .nr(16)
19282 .kr(16)
19283 .sr(1)
19284 .m(4)
19285 .n(n)
19286 .k(16)
19287 .iterations(1)
19288 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19289 }
19290 }
19291
19292 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_lt_16) {
19293 TEST_REQUIRES_ARM_NEON;
19294 for (size_t k = 1; k < 16; k++) {
19295 GemmMicrokernelTester()
19296 .mr(4)
19297 .nr(16)
19298 .kr(16)
19299 .sr(1)
19300 .m(4)
19301 .n(16)
19302 .k(k)
19303 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19304 }
19305 }
19306
19307 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_lt_16_subtile) {
19308 TEST_REQUIRES_ARM_NEON;
19309 for (size_t k = 1; k < 16; k++) {
19310 for (uint32_t n = 1; n <= 16; n++) {
19311 for (uint32_t m = 1; m <= 4; m++) {
19312 GemmMicrokernelTester()
19313 .mr(4)
19314 .nr(16)
19315 .kr(16)
19316 .sr(1)
19317 .m(m)
19318 .n(n)
19319 .k(k)
19320 .iterations(1)
19321 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19322 }
19323 }
19324 }
19325 }
19326
19327 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_gt_16) {
19328 TEST_REQUIRES_ARM_NEON;
19329 for (size_t k = 17; k < 32; k++) {
19330 GemmMicrokernelTester()
19331 .mr(4)
19332 .nr(16)
19333 .kr(16)
19334 .sr(1)
19335 .m(4)
19336 .n(16)
19337 .k(k)
19338 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19339 }
19340 }
19341
19342 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_gt_16_subtile) {
19343 TEST_REQUIRES_ARM_NEON;
19344 for (size_t k = 17; k < 32; k++) {
19345 for (uint32_t n = 1; n <= 16; n++) {
19346 for (uint32_t m = 1; m <= 4; m++) {
19347 GemmMicrokernelTester()
19348 .mr(4)
19349 .nr(16)
19350 .kr(16)
19351 .sr(1)
19352 .m(m)
19353 .n(n)
19354 .k(k)
19355 .iterations(1)
19356 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19357 }
19358 }
19359 }
19360 }
19361
19362 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_div_16) {
19363 TEST_REQUIRES_ARM_NEON;
19364 for (size_t k = 32; k <= 160; k += 16) {
19365 GemmMicrokernelTester()
19366 .mr(4)
19367 .nr(16)
19368 .kr(16)
19369 .sr(1)
19370 .m(4)
19371 .n(16)
19372 .k(k)
19373 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19374 }
19375 }
19376
19377 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, k_div_16_subtile) {
19378 TEST_REQUIRES_ARM_NEON;
19379 for (size_t k = 32; k <= 160; k += 16) {
19380 for (uint32_t n = 1; n <= 16; n++) {
19381 for (uint32_t m = 1; m <= 4; m++) {
19382 GemmMicrokernelTester()
19383 .mr(4)
19384 .nr(16)
19385 .kr(16)
19386 .sr(1)
19387 .m(m)
19388 .n(n)
19389 .k(k)
19390 .iterations(1)
19391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19392 }
19393 }
19394 }
19395 }
19396
19397 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16) {
19398 TEST_REQUIRES_ARM_NEON;
19399 for (uint32_t n = 17; n < 32; n++) {
19400 for (size_t k = 1; k <= 80; k += 17) {
19401 GemmMicrokernelTester()
19402 .mr(4)
19403 .nr(16)
19404 .kr(16)
19405 .sr(1)
19406 .m(4)
19407 .n(n)
19408 .k(k)
19409 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19410 }
19411 }
19412 }
19413
19414 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16_strided_cn) {
19415 TEST_REQUIRES_ARM_NEON;
19416 for (uint32_t n = 17; n < 32; n++) {
19417 for (size_t k = 1; k <= 80; k += 17) {
19418 GemmMicrokernelTester()
19419 .mr(4)
19420 .nr(16)
19421 .kr(16)
19422 .sr(1)
19423 .m(4)
19424 .n(n)
19425 .k(k)
19426 .cn_stride(19)
19427 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19428 }
19429 }
19430 }
19431
19432 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16_subtile) {
19433 TEST_REQUIRES_ARM_NEON;
19434 for (uint32_t n = 17; n < 32; n++) {
19435 for (size_t k = 1; k <= 80; k += 17) {
19436 for (uint32_t m = 1; m <= 4; m++) {
19437 GemmMicrokernelTester()
19438 .mr(4)
19439 .nr(16)
19440 .kr(16)
19441 .sr(1)
19442 .m(m)
19443 .n(n)
19444 .k(k)
19445 .iterations(1)
19446 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19447 }
19448 }
19449 }
19450 }
19451
19452 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16) {
19453 TEST_REQUIRES_ARM_NEON;
19454 for (uint32_t n = 32; n <= 48; n += 16) {
19455 for (size_t k = 1; k <= 80; k += 17) {
19456 GemmMicrokernelTester()
19457 .mr(4)
19458 .nr(16)
19459 .kr(16)
19460 .sr(1)
19461 .m(4)
19462 .n(n)
19463 .k(k)
19464 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19465 }
19466 }
19467 }
19468
19469 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16_strided_cn) {
19470 TEST_REQUIRES_ARM_NEON;
19471 for (uint32_t n = 32; n <= 48; n += 16) {
19472 for (size_t k = 1; k <= 80; k += 17) {
19473 GemmMicrokernelTester()
19474 .mr(4)
19475 .nr(16)
19476 .kr(16)
19477 .sr(1)
19478 .m(4)
19479 .n(n)
19480 .k(k)
19481 .cn_stride(19)
19482 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19483 }
19484 }
19485 }
19486
19487 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16_subtile) {
19488 TEST_REQUIRES_ARM_NEON;
19489 for (uint32_t n = 32; n <= 48; n += 16) {
19490 for (size_t k = 1; k <= 80; k += 17) {
19491 for (uint32_t m = 1; m <= 4; m++) {
19492 GemmMicrokernelTester()
19493 .mr(4)
19494 .nr(16)
19495 .kr(16)
19496 .sr(1)
19497 .m(m)
19498 .n(n)
19499 .k(k)
19500 .iterations(1)
19501 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19502 }
19503 }
19504 }
19505 }
19506
19507 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, small_kernel) {
19508 TEST_REQUIRES_ARM_NEON;
19509 for (size_t k = 1; k <= 80; k += 17) {
19510 GemmMicrokernelTester()
19511 .mr(4)
19512 .nr(16)
19513 .kr(16)
19514 .sr(1)
19515 .m(4)
19516 .n(16)
19517 .k(k)
19518 .ks(3)
19519 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19520 }
19521 }
19522
19523 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, small_kernel_subtile) {
19524 TEST_REQUIRES_ARM_NEON;
19525 for (size_t k = 1; k <= 80; k += 17) {
19526 for (uint32_t n = 1; n <= 16; n++) {
19527 for (uint32_t m = 1; m <= 4; m++) {
19528 GemmMicrokernelTester()
19529 .mr(4)
19530 .nr(16)
19531 .kr(16)
19532 .sr(1)
19533 .m(m)
19534 .n(n)
19535 .k(k)
19536 .ks(3)
19537 .iterations(1)
19538 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19539 }
19540 }
19541 }
19542 }
19543
19544 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_gt_16_small_kernel) {
19545 TEST_REQUIRES_ARM_NEON;
19546 for (uint32_t n = 17; n < 32; n++) {
19547 for (size_t k = 1; k <= 80; k += 17) {
19548 GemmMicrokernelTester()
19549 .mr(4)
19550 .nr(16)
19551 .kr(16)
19552 .sr(1)
19553 .m(4)
19554 .n(n)
19555 .k(k)
19556 .ks(3)
19557 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19558 }
19559 }
19560 }
19561
19562 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, n_div_16_small_kernel) {
19563 TEST_REQUIRES_ARM_NEON;
19564 for (uint32_t n = 32; n <= 48; n += 16) {
19565 for (size_t k = 1; k <= 80; k += 17) {
19566 GemmMicrokernelTester()
19567 .mr(4)
19568 .nr(16)
19569 .kr(16)
19570 .sr(1)
19571 .m(4)
19572 .n(n)
19573 .k(k)
19574 .ks(3)
19575 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19576 }
19577 }
19578 }
19579
19580 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, strided_cm_subtile) {
19581 TEST_REQUIRES_ARM_NEON;
19582 for (size_t k = 1; k <= 80; k += 17) {
19583 for (uint32_t n = 1; n <= 16; n++) {
19584 for (uint32_t m = 1; m <= 4; m++) {
19585 GemmMicrokernelTester()
19586 .mr(4)
19587 .nr(16)
19588 .kr(16)
19589 .sr(1)
19590 .m(m)
19591 .n(n)
19592 .k(k)
19593 .cm_stride(19)
19594 .iterations(1)
19595 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19596 }
19597 }
19598 }
19599 }
19600
19601 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, a_offset) {
19602 TEST_REQUIRES_ARM_NEON;
19603 for (size_t k = 1; k <= 80; k += 17) {
19604 GemmMicrokernelTester()
19605 .mr(4)
19606 .nr(16)
19607 .kr(16)
19608 .sr(1)
19609 .m(4)
19610 .n(16)
19611 .k(k)
19612 .ks(3)
19613 .a_offset(331)
19614 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19615 }
19616 }
19617
19618 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, zero) {
19619 TEST_REQUIRES_ARM_NEON;
19620 for (size_t k = 1; k <= 80; k += 17) {
19621 for (uint32_t mz = 0; mz < 4; mz++) {
19622 GemmMicrokernelTester()
19623 .mr(4)
19624 .nr(16)
19625 .kr(16)
19626 .sr(1)
19627 .m(4)
19628 .n(16)
19629 .k(k)
19630 .ks(3)
19631 .a_offset(331)
19632 .zero_index(mz)
19633 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19634 }
19635 }
19636 }
19637
19638 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, qmin) {
19639 TEST_REQUIRES_ARM_NEON;
19640 GemmMicrokernelTester()
19641 .mr(4)
19642 .nr(16)
19643 .kr(16)
19644 .sr(1)
19645 .m(4)
19646 .n(16)
19647 .k(16)
19648 .qmin(128)
19649 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19650 }
19651
19652 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, qmax) {
19653 TEST_REQUIRES_ARM_NEON;
19654 GemmMicrokernelTester()
19655 .mr(4)
19656 .nr(16)
19657 .kr(16)
19658 .sr(1)
19659 .m(4)
19660 .n(16)
19661 .k(16)
19662 .qmax(128)
19663 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19664 }
19665
19666 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16C16__NEON_MLAL, strided_cm) {
19667 TEST_REQUIRES_ARM_NEON;
19668 GemmMicrokernelTester()
19669 .mr(4)
19670 .nr(16)
19671 .kr(16)
19672 .sr(1)
19673 .m(4)
19674 .n(16)
19675 .k(16)
19676 .cm_stride(19)
19677 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16c16__neon_mlal, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19678 }
19679#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
19680
19681
19682#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
19683 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8) {
19684 TEST_REQUIRES_ARM_NEON_DOT;
19685 GemmMicrokernelTester()
19686 .mr(6)
19687 .nr(16)
19688 .kr(4)
19689 .sr(1)
19690 .m(6)
19691 .n(16)
19692 .k(8)
19693 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19694 }
19695
19696 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cn) {
19697 TEST_REQUIRES_ARM_NEON_DOT;
19698 GemmMicrokernelTester()
19699 .mr(6)
19700 .nr(16)
19701 .kr(4)
19702 .sr(1)
19703 .m(6)
19704 .n(16)
19705 .k(8)
19706 .cn_stride(19)
19707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19708 }
19709
19710 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile) {
19711 TEST_REQUIRES_ARM_NEON_DOT;
19712 for (uint32_t n = 1; n <= 16; n++) {
19713 for (uint32_t m = 1; m <= 6; m++) {
19714 GemmMicrokernelTester()
19715 .mr(6)
19716 .nr(16)
19717 .kr(4)
19718 .sr(1)
19719 .m(m)
19720 .n(n)
19721 .k(8)
19722 .iterations(1)
19723 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19724 }
19725 }
19726 }
19727
19728 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile_m) {
19729 TEST_REQUIRES_ARM_NEON_DOT;
19730 for (uint32_t m = 1; m <= 6; m++) {
19731 GemmMicrokernelTester()
19732 .mr(6)
19733 .nr(16)
19734 .kr(4)
19735 .sr(1)
19736 .m(m)
19737 .n(16)
19738 .k(8)
19739 .iterations(1)
19740 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19741 }
19742 }
19743
19744 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_eq_8_subtile_n) {
19745 TEST_REQUIRES_ARM_NEON_DOT;
19746 for (uint32_t n = 1; n <= 16; n++) {
19747 GemmMicrokernelTester()
19748 .mr(6)
19749 .nr(16)
19750 .kr(4)
19751 .sr(1)
19752 .m(6)
19753 .n(n)
19754 .k(8)
19755 .iterations(1)
19756 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19757 }
19758 }
19759
19760 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8) {
19761 TEST_REQUIRES_ARM_NEON_DOT;
19762 for (size_t k = 1; k < 8; k++) {
19763 GemmMicrokernelTester()
19764 .mr(6)
19765 .nr(16)
19766 .kr(4)
19767 .sr(1)
19768 .m(6)
19769 .n(16)
19770 .k(k)
19771 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19772 }
19773 }
19774
19775 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_lt_8_subtile) {
19776 TEST_REQUIRES_ARM_NEON_DOT;
19777 for (size_t k = 1; k < 8; k++) {
19778 for (uint32_t n = 1; n <= 16; n++) {
19779 for (uint32_t m = 1; m <= 6; m++) {
19780 GemmMicrokernelTester()
19781 .mr(6)
19782 .nr(16)
19783 .kr(4)
19784 .sr(1)
19785 .m(m)
19786 .n(n)
19787 .k(k)
19788 .iterations(1)
19789 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19790 }
19791 }
19792 }
19793 }
19794
19795 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8) {
19796 TEST_REQUIRES_ARM_NEON_DOT;
19797 for (size_t k = 9; k < 16; k++) {
19798 GemmMicrokernelTester()
19799 .mr(6)
19800 .nr(16)
19801 .kr(4)
19802 .sr(1)
19803 .m(6)
19804 .n(16)
19805 .k(k)
19806 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19807 }
19808 }
19809
19810 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_gt_8_subtile) {
19811 TEST_REQUIRES_ARM_NEON_DOT;
19812 for (size_t k = 9; k < 16; k++) {
19813 for (uint32_t n = 1; n <= 16; n++) {
19814 for (uint32_t m = 1; m <= 6; m++) {
19815 GemmMicrokernelTester()
19816 .mr(6)
19817 .nr(16)
19818 .kr(4)
19819 .sr(1)
19820 .m(m)
19821 .n(n)
19822 .k(k)
19823 .iterations(1)
19824 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19825 }
19826 }
19827 }
19828 }
19829
19830 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8) {
19831 TEST_REQUIRES_ARM_NEON_DOT;
19832 for (size_t k = 16; k <= 80; k += 8) {
19833 GemmMicrokernelTester()
19834 .mr(6)
19835 .nr(16)
19836 .kr(4)
19837 .sr(1)
19838 .m(6)
19839 .n(16)
19840 .k(k)
19841 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19842 }
19843 }
19844
19845 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, k_div_8_subtile) {
19846 TEST_REQUIRES_ARM_NEON_DOT;
19847 for (size_t k = 16; k <= 80; k += 8) {
19848 for (uint32_t n = 1; n <= 16; n++) {
19849 for (uint32_t m = 1; m <= 6; m++) {
19850 GemmMicrokernelTester()
19851 .mr(6)
19852 .nr(16)
19853 .kr(4)
19854 .sr(1)
19855 .m(m)
19856 .n(n)
19857 .k(k)
19858 .iterations(1)
19859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19860 }
19861 }
19862 }
19863 }
19864
19865 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16) {
19866 TEST_REQUIRES_ARM_NEON_DOT;
19867 for (uint32_t n = 17; n < 32; n++) {
19868 for (size_t k = 1; k <= 40; k += 9) {
19869 GemmMicrokernelTester()
19870 .mr(6)
19871 .nr(16)
19872 .kr(4)
19873 .sr(1)
19874 .m(6)
19875 .n(n)
19876 .k(k)
19877 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19878 }
19879 }
19880 }
19881
19882 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_strided_cn) {
19883 TEST_REQUIRES_ARM_NEON_DOT;
19884 for (uint32_t n = 17; n < 32; n++) {
19885 for (size_t k = 1; k <= 40; k += 9) {
19886 GemmMicrokernelTester()
19887 .mr(6)
19888 .nr(16)
19889 .kr(4)
19890 .sr(1)
19891 .m(6)
19892 .n(n)
19893 .k(k)
19894 .cn_stride(19)
19895 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19896 }
19897 }
19898 }
19899
19900 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_subtile) {
19901 TEST_REQUIRES_ARM_NEON_DOT;
19902 for (uint32_t n = 17; n < 32; n++) {
19903 for (size_t k = 1; k <= 40; k += 9) {
19904 for (uint32_t m = 1; m <= 6; m++) {
19905 GemmMicrokernelTester()
19906 .mr(6)
19907 .nr(16)
19908 .kr(4)
19909 .sr(1)
19910 .m(m)
19911 .n(n)
19912 .k(k)
19913 .iterations(1)
19914 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19915 }
19916 }
19917 }
19918 }
19919
19920 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16) {
19921 TEST_REQUIRES_ARM_NEON_DOT;
19922 for (uint32_t n = 32; n <= 48; n += 16) {
19923 for (size_t k = 1; k <= 40; k += 9) {
19924 GemmMicrokernelTester()
19925 .mr(6)
19926 .nr(16)
19927 .kr(4)
19928 .sr(1)
19929 .m(6)
19930 .n(n)
19931 .k(k)
19932 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19933 }
19934 }
19935 }
19936
19937 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_strided_cn) {
19938 TEST_REQUIRES_ARM_NEON_DOT;
19939 for (uint32_t n = 32; n <= 48; n += 16) {
19940 for (size_t k = 1; k <= 40; k += 9) {
19941 GemmMicrokernelTester()
19942 .mr(6)
19943 .nr(16)
19944 .kr(4)
19945 .sr(1)
19946 .m(6)
19947 .n(n)
19948 .k(k)
19949 .cn_stride(19)
19950 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19951 }
19952 }
19953 }
19954
19955 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_subtile) {
19956 TEST_REQUIRES_ARM_NEON_DOT;
19957 for (uint32_t n = 32; n <= 48; n += 16) {
19958 for (size_t k = 1; k <= 40; k += 9) {
19959 for (uint32_t m = 1; m <= 6; m++) {
19960 GemmMicrokernelTester()
19961 .mr(6)
19962 .nr(16)
19963 .kr(4)
19964 .sr(1)
19965 .m(m)
19966 .n(n)
19967 .k(k)
19968 .iterations(1)
19969 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19970 }
19971 }
19972 }
19973 }
19974
19975 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, small_kernel) {
19976 TEST_REQUIRES_ARM_NEON_DOT;
19977 for (size_t k = 1; k <= 40; k += 9) {
19978 GemmMicrokernelTester()
19979 .mr(6)
19980 .nr(16)
19981 .kr(4)
19982 .sr(1)
19983 .m(6)
19984 .n(16)
19985 .k(k)
19986 .ks(3)
19987 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
19988 }
19989 }
19990
19991 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, small_kernel_subtile) {
19992 TEST_REQUIRES_ARM_NEON_DOT;
19993 for (size_t k = 1; k <= 40; k += 9) {
19994 for (uint32_t n = 1; n <= 16; n++) {
19995 for (uint32_t m = 1; m <= 6; m++) {
19996 GemmMicrokernelTester()
19997 .mr(6)
19998 .nr(16)
19999 .kr(4)
20000 .sr(1)
20001 .m(m)
20002 .n(n)
20003 .k(k)
20004 .ks(3)
20005 .iterations(1)
20006 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20007 }
20008 }
20009 }
20010 }
20011
20012 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_gt_16_small_kernel) {
20013 TEST_REQUIRES_ARM_NEON_DOT;
20014 for (uint32_t n = 17; n < 32; n++) {
20015 for (size_t k = 1; k <= 40; k += 9) {
20016 GemmMicrokernelTester()
20017 .mr(6)
20018 .nr(16)
20019 .kr(4)
20020 .sr(1)
20021 .m(6)
20022 .n(n)
20023 .k(k)
20024 .ks(3)
20025 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20026 }
20027 }
20028 }
20029
20030 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, n_div_16_small_kernel) {
20031 TEST_REQUIRES_ARM_NEON_DOT;
20032 for (uint32_t n = 32; n <= 48; n += 16) {
20033 for (size_t k = 1; k <= 40; k += 9) {
20034 GemmMicrokernelTester()
20035 .mr(6)
20036 .nr(16)
20037 .kr(4)
20038 .sr(1)
20039 .m(6)
20040 .n(n)
20041 .k(k)
20042 .ks(3)
20043 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20044 }
20045 }
20046 }
20047
20048 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cm_subtile) {
20049 TEST_REQUIRES_ARM_NEON_DOT;
20050 for (size_t k = 1; k <= 40; k += 9) {
20051 for (uint32_t n = 1; n <= 16; n++) {
20052 for (uint32_t m = 1; m <= 6; m++) {
20053 GemmMicrokernelTester()
20054 .mr(6)
20055 .nr(16)
20056 .kr(4)
20057 .sr(1)
20058 .m(m)
20059 .n(n)
20060 .k(k)
20061 .cm_stride(19)
20062 .iterations(1)
20063 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20064 }
20065 }
20066 }
20067 }
20068
20069 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, a_offset) {
20070 TEST_REQUIRES_ARM_NEON_DOT;
20071 for (size_t k = 1; k <= 40; k += 9) {
20072 GemmMicrokernelTester()
20073 .mr(6)
20074 .nr(16)
20075 .kr(4)
20076 .sr(1)
20077 .m(6)
20078 .n(16)
20079 .k(k)
20080 .ks(3)
20081 .a_offset(251)
20082 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20083 }
20084 }
20085
20086 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, zero) {
20087 TEST_REQUIRES_ARM_NEON_DOT;
20088 for (size_t k = 1; k <= 40; k += 9) {
20089 for (uint32_t mz = 0; mz < 6; mz++) {
20090 GemmMicrokernelTester()
20091 .mr(6)
20092 .nr(16)
20093 .kr(4)
20094 .sr(1)
20095 .m(6)
20096 .n(16)
20097 .k(k)
20098 .ks(3)
20099 .a_offset(251)
20100 .zero_index(mz)
20101 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20102 }
20103 }
20104 }
20105
20106 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, qmin) {
20107 TEST_REQUIRES_ARM_NEON_DOT;
20108 GemmMicrokernelTester()
20109 .mr(6)
20110 .nr(16)
20111 .kr(4)
20112 .sr(1)
20113 .m(6)
20114 .n(16)
20115 .k(8)
20116 .qmin(128)
20117 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20118 }
20119
20120 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, qmax) {
20121 TEST_REQUIRES_ARM_NEON_DOT;
20122 GemmMicrokernelTester()
20123 .mr(6)
20124 .nr(16)
20125 .kr(4)
20126 .sr(1)
20127 .m(6)
20128 .n(16)
20129 .k(8)
20130 .qmax(128)
20131 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20132 }
20133
20134 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16C4__NEONDOT, strided_cm) {
20135 TEST_REQUIRES_ARM_NEON_DOT;
20136 GemmMicrokernelTester()
20137 .mr(6)
20138 .nr(16)
20139 .kr(4)
20140 .sr(1)
20141 .m(6)
20142 .n(16)
20143 .k(8)
20144 .cm_stride(19)
20145 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16c4__neondot, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20146 }
20147#endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
20148
20149
20150#if XNN_ARCH_ARM || XNN_ARCH_ARM64
20151 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8) {
20152 TEST_REQUIRES_ARM_NEON;
20153 GemmMicrokernelTester()
20154 .mr(6)
20155 .nr(8)
20156 .kr(1)
20157 .sr(1)
20158 .m(6)
20159 .n(8)
20160 .k(8)
20161 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20162 }
20163
20164 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, strided_cn) {
20165 TEST_REQUIRES_ARM_NEON;
20166 GemmMicrokernelTester()
20167 .mr(6)
20168 .nr(8)
20169 .kr(1)
20170 .sr(1)
20171 .m(6)
20172 .n(8)
20173 .k(8)
20174 .cn_stride(11)
20175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20176 }
20177
20178 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8_subtile) {
20179 TEST_REQUIRES_ARM_NEON;
20180 for (uint32_t n = 1; n <= 8; n++) {
20181 for (uint32_t m = 1; m <= 6; m++) {
20182 GemmMicrokernelTester()
20183 .mr(6)
20184 .nr(8)
20185 .kr(1)
20186 .sr(1)
20187 .m(m)
20188 .n(n)
20189 .k(8)
20190 .iterations(1)
20191 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20192 }
20193 }
20194 }
20195
20196 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8_subtile_m) {
20197 TEST_REQUIRES_ARM_NEON;
20198 for (uint32_t m = 1; m <= 6; m++) {
20199 GemmMicrokernelTester()
20200 .mr(6)
20201 .nr(8)
20202 .kr(1)
20203 .sr(1)
20204 .m(m)
20205 .n(8)
20206 .k(8)
20207 .iterations(1)
20208 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20209 }
20210 }
20211
20212 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_eq_8_subtile_n) {
20213 TEST_REQUIRES_ARM_NEON;
20214 for (uint32_t n = 1; n <= 8; n++) {
20215 GemmMicrokernelTester()
20216 .mr(6)
20217 .nr(8)
20218 .kr(1)
20219 .sr(1)
20220 .m(6)
20221 .n(n)
20222 .k(8)
20223 .iterations(1)
20224 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20225 }
20226 }
20227
20228 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_lt_8) {
20229 TEST_REQUIRES_ARM_NEON;
20230 for (size_t k = 1; k < 8; k++) {
20231 GemmMicrokernelTester()
20232 .mr(6)
20233 .nr(8)
20234 .kr(1)
20235 .sr(1)
20236 .m(6)
20237 .n(8)
20238 .k(k)
20239 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20240 }
20241 }
20242
20243 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_lt_8_subtile) {
20244 TEST_REQUIRES_ARM_NEON;
20245 for (size_t k = 1; k < 8; k++) {
20246 for (uint32_t n = 1; n <= 8; n++) {
20247 for (uint32_t m = 1; m <= 6; m++) {
20248 GemmMicrokernelTester()
20249 .mr(6)
20250 .nr(8)
20251 .kr(1)
20252 .sr(1)
20253 .m(m)
20254 .n(n)
20255 .k(k)
20256 .iterations(1)
20257 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20258 }
20259 }
20260 }
20261 }
20262
20263 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_gt_8) {
20264 TEST_REQUIRES_ARM_NEON;
20265 for (size_t k = 9; k < 16; k++) {
20266 GemmMicrokernelTester()
20267 .mr(6)
20268 .nr(8)
20269 .kr(1)
20270 .sr(1)
20271 .m(6)
20272 .n(8)
20273 .k(k)
20274 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20275 }
20276 }
20277
20278 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_gt_8_subtile) {
20279 TEST_REQUIRES_ARM_NEON;
20280 for (size_t k = 9; k < 16; k++) {
20281 for (uint32_t n = 1; n <= 8; n++) {
20282 for (uint32_t m = 1; m <= 6; m++) {
20283 GemmMicrokernelTester()
20284 .mr(6)
20285 .nr(8)
20286 .kr(1)
20287 .sr(1)
20288 .m(m)
20289 .n(n)
20290 .k(k)
20291 .iterations(1)
20292 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20293 }
20294 }
20295 }
20296 }
20297
20298 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_div_8) {
20299 TEST_REQUIRES_ARM_NEON;
20300 for (size_t k = 16; k <= 80; k += 8) {
20301 GemmMicrokernelTester()
20302 .mr(6)
20303 .nr(8)
20304 .kr(1)
20305 .sr(1)
20306 .m(6)
20307 .n(8)
20308 .k(k)
20309 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20310 }
20311 }
20312
20313 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, k_div_8_subtile) {
20314 TEST_REQUIRES_ARM_NEON;
20315 for (size_t k = 16; k <= 80; k += 8) {
20316 for (uint32_t n = 1; n <= 8; n++) {
20317 for (uint32_t m = 1; m <= 6; m++) {
20318 GemmMicrokernelTester()
20319 .mr(6)
20320 .nr(8)
20321 .kr(1)
20322 .sr(1)
20323 .m(m)
20324 .n(n)
20325 .k(k)
20326 .iterations(1)
20327 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20328 }
20329 }
20330 }
20331 }
20332
20333 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8) {
20334 TEST_REQUIRES_ARM_NEON;
20335 for (uint32_t n = 9; n < 16; n++) {
20336 for (size_t k = 1; k <= 40; k += 9) {
20337 GemmMicrokernelTester()
20338 .mr(6)
20339 .nr(8)
20340 .kr(1)
20341 .sr(1)
20342 .m(6)
20343 .n(n)
20344 .k(k)
20345 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20346 }
20347 }
20348 }
20349
20350 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8_strided_cn) {
20351 TEST_REQUIRES_ARM_NEON;
20352 for (uint32_t n = 9; n < 16; n++) {
20353 for (size_t k = 1; k <= 40; k += 9) {
20354 GemmMicrokernelTester()
20355 .mr(6)
20356 .nr(8)
20357 .kr(1)
20358 .sr(1)
20359 .m(6)
20360 .n(n)
20361 .k(k)
20362 .cn_stride(11)
20363 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20364 }
20365 }
20366 }
20367
20368 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8_subtile) {
20369 TEST_REQUIRES_ARM_NEON;
20370 for (uint32_t n = 9; n < 16; n++) {
20371 for (size_t k = 1; k <= 40; k += 9) {
20372 for (uint32_t m = 1; m <= 6; m++) {
20373 GemmMicrokernelTester()
20374 .mr(6)
20375 .nr(8)
20376 .kr(1)
20377 .sr(1)
20378 .m(m)
20379 .n(n)
20380 .k(k)
20381 .iterations(1)
20382 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20383 }
20384 }
20385 }
20386 }
20387
20388 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8) {
20389 TEST_REQUIRES_ARM_NEON;
20390 for (uint32_t n = 16; n <= 24; n += 8) {
20391 for (size_t k = 1; k <= 40; k += 9) {
20392 GemmMicrokernelTester()
20393 .mr(6)
20394 .nr(8)
20395 .kr(1)
20396 .sr(1)
20397 .m(6)
20398 .n(n)
20399 .k(k)
20400 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20401 }
20402 }
20403 }
20404
20405 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8_strided_cn) {
20406 TEST_REQUIRES_ARM_NEON;
20407 for (uint32_t n = 16; n <= 24; n += 8) {
20408 for (size_t k = 1; k <= 40; k += 9) {
20409 GemmMicrokernelTester()
20410 .mr(6)
20411 .nr(8)
20412 .kr(1)
20413 .sr(1)
20414 .m(6)
20415 .n(n)
20416 .k(k)
20417 .cn_stride(11)
20418 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20419 }
20420 }
20421 }
20422
20423 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8_subtile) {
20424 TEST_REQUIRES_ARM_NEON;
20425 for (uint32_t n = 16; n <= 24; n += 8) {
20426 for (size_t k = 1; k <= 40; k += 9) {
20427 for (uint32_t m = 1; m <= 6; m++) {
20428 GemmMicrokernelTester()
20429 .mr(6)
20430 .nr(8)
20431 .kr(1)
20432 .sr(1)
20433 .m(m)
20434 .n(n)
20435 .k(k)
20436 .iterations(1)
20437 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20438 }
20439 }
20440 }
20441 }
20442
20443 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, small_kernel) {
20444 TEST_REQUIRES_ARM_NEON;
20445 for (size_t k = 1; k <= 40; k += 9) {
20446 GemmMicrokernelTester()
20447 .mr(6)
20448 .nr(8)
20449 .kr(1)
20450 .sr(1)
20451 .m(6)
20452 .n(8)
20453 .k(k)
20454 .ks(3)
20455 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20456 }
20457 }
20458
20459 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, small_kernel_subtile) {
20460 TEST_REQUIRES_ARM_NEON;
20461 for (size_t k = 1; k <= 40; k += 9) {
20462 for (uint32_t n = 1; n <= 8; n++) {
20463 for (uint32_t m = 1; m <= 6; m++) {
20464 GemmMicrokernelTester()
20465 .mr(6)
20466 .nr(8)
20467 .kr(1)
20468 .sr(1)
20469 .m(m)
20470 .n(n)
20471 .k(k)
20472 .ks(3)
20473 .iterations(1)
20474 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20475 }
20476 }
20477 }
20478 }
20479
20480 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_gt_8_small_kernel) {
20481 TEST_REQUIRES_ARM_NEON;
20482 for (uint32_t n = 9; n < 16; n++) {
20483 for (size_t k = 1; k <= 40; k += 9) {
20484 GemmMicrokernelTester()
20485 .mr(6)
20486 .nr(8)
20487 .kr(1)
20488 .sr(1)
20489 .m(6)
20490 .n(n)
20491 .k(k)
20492 .ks(3)
20493 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20494 }
20495 }
20496 }
20497
20498 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, n_div_8_small_kernel) {
20499 TEST_REQUIRES_ARM_NEON;
20500 for (uint32_t n = 16; n <= 24; n += 8) {
20501 for (size_t k = 1; k <= 40; k += 9) {
20502 GemmMicrokernelTester()
20503 .mr(6)
20504 .nr(8)
20505 .kr(1)
20506 .sr(1)
20507 .m(6)
20508 .n(n)
20509 .k(k)
20510 .ks(3)
20511 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20512 }
20513 }
20514 }
20515
20516 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, strided_cm_subtile) {
20517 TEST_REQUIRES_ARM_NEON;
20518 for (size_t k = 1; k <= 40; k += 9) {
20519 for (uint32_t n = 1; n <= 8; n++) {
20520 for (uint32_t m = 1; m <= 6; m++) {
20521 GemmMicrokernelTester()
20522 .mr(6)
20523 .nr(8)
20524 .kr(1)
20525 .sr(1)
20526 .m(m)
20527 .n(n)
20528 .k(k)
20529 .cm_stride(11)
20530 .iterations(1)
20531 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20532 }
20533 }
20534 }
20535 }
20536
20537 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, a_offset) {
20538 TEST_REQUIRES_ARM_NEON;
20539 for (size_t k = 1; k <= 40; k += 9) {
20540 GemmMicrokernelTester()
20541 .mr(6)
20542 .nr(8)
20543 .kr(1)
20544 .sr(1)
20545 .m(6)
20546 .n(8)
20547 .k(k)
20548 .ks(3)
20549 .a_offset(251)
20550 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20551 }
20552 }
20553
20554 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, zero) {
20555 TEST_REQUIRES_ARM_NEON;
20556 for (size_t k = 1; k <= 40; k += 9) {
20557 for (uint32_t mz = 0; mz < 6; mz++) {
20558 GemmMicrokernelTester()
20559 .mr(6)
20560 .nr(8)
20561 .kr(1)
20562 .sr(1)
20563 .m(6)
20564 .n(8)
20565 .k(k)
20566 .ks(3)
20567 .a_offset(251)
20568 .zero_index(mz)
20569 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20570 }
20571 }
20572 }
20573
20574 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, qmin) {
20575 TEST_REQUIRES_ARM_NEON;
20576 GemmMicrokernelTester()
20577 .mr(6)
20578 .nr(8)
20579 .kr(1)
20580 .sr(1)
20581 .m(6)
20582 .n(8)
20583 .k(8)
20584 .qmin(128)
20585 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20586 }
20587
20588 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, qmax) {
20589 TEST_REQUIRES_ARM_NEON;
20590 GemmMicrokernelTester()
20591 .mr(6)
20592 .nr(8)
20593 .kr(1)
20594 .sr(1)
20595 .m(6)
20596 .n(8)
20597 .k(8)
20598 .qmax(128)
20599 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20600 }
20601
20602 TEST(QS8_IGEMM_MINMAX_RNDNU_6X8__NEON_MLAL_LANE, strided_cm) {
20603 TEST_REQUIRES_ARM_NEON;
20604 GemmMicrokernelTester()
20605 .mr(6)
20606 .nr(8)
20607 .kr(1)
20608 .sr(1)
20609 .m(6)
20610 .n(8)
20611 .k(8)
20612 .cm_stride(11)
20613 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x8__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20614 }
20615#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
20616
20617
20618#if XNN_ARCH_ARM || XNN_ARCH_ARM64
20619 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_eq_8) {
20620 TEST_REQUIRES_ARM_NEON;
20621 GemmMicrokernelTester()
20622 .mr(4)
20623 .nr(16)
20624 .kr(1)
20625 .sr(1)
20626 .m(4)
20627 .n(16)
20628 .k(8)
20629 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20630 }
20631
20632 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, strided_cn) {
20633 TEST_REQUIRES_ARM_NEON;
20634 GemmMicrokernelTester()
20635 .mr(4)
20636 .nr(16)
20637 .kr(1)
20638 .sr(1)
20639 .m(4)
20640 .n(16)
20641 .k(8)
20642 .cn_stride(19)
20643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20644 }
20645
20646 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_eq_8_subtile) {
20647 TEST_REQUIRES_ARM_NEON;
20648 for (uint32_t n = 1; n <= 16; n++) {
20649 for (uint32_t m = 1; m <= 4; m++) {
20650 GemmMicrokernelTester()
20651 .mr(4)
20652 .nr(16)
20653 .kr(1)
20654 .sr(1)
20655 .m(m)
20656 .n(n)
20657 .k(8)
20658 .iterations(1)
20659 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20660 }
20661 }
20662 }
20663
20664 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_eq_8_subtile_m) {
20665 TEST_REQUIRES_ARM_NEON;
20666 for (uint32_t m = 1; m <= 4; m++) {
20667 GemmMicrokernelTester()
20668 .mr(4)
20669 .nr(16)
20670 .kr(1)
20671 .sr(1)
20672 .m(m)
20673 .n(16)
20674 .k(8)
20675 .iterations(1)
20676 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20677 }
20678 }
20679
20680 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_eq_8_subtile_n) {
20681 TEST_REQUIRES_ARM_NEON;
20682 for (uint32_t n = 1; n <= 16; n++) {
20683 GemmMicrokernelTester()
20684 .mr(4)
20685 .nr(16)
20686 .kr(1)
20687 .sr(1)
20688 .m(4)
20689 .n(n)
20690 .k(8)
20691 .iterations(1)
20692 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20693 }
20694 }
20695
20696 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_lt_8) {
20697 TEST_REQUIRES_ARM_NEON;
20698 for (size_t k = 1; k < 8; k++) {
20699 GemmMicrokernelTester()
20700 .mr(4)
20701 .nr(16)
20702 .kr(1)
20703 .sr(1)
20704 .m(4)
20705 .n(16)
20706 .k(k)
20707 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20708 }
20709 }
20710
20711 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_lt_8_subtile) {
20712 TEST_REQUIRES_ARM_NEON;
20713 for (size_t k = 1; k < 8; k++) {
20714 for (uint32_t n = 1; n <= 16; n++) {
20715 for (uint32_t m = 1; m <= 4; m++) {
20716 GemmMicrokernelTester()
20717 .mr(4)
20718 .nr(16)
20719 .kr(1)
20720 .sr(1)
20721 .m(m)
20722 .n(n)
20723 .k(k)
20724 .iterations(1)
20725 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20726 }
20727 }
20728 }
20729 }
20730
20731 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_gt_8) {
20732 TEST_REQUIRES_ARM_NEON;
20733 for (size_t k = 9; k < 16; k++) {
20734 GemmMicrokernelTester()
20735 .mr(4)
20736 .nr(16)
20737 .kr(1)
20738 .sr(1)
20739 .m(4)
20740 .n(16)
20741 .k(k)
20742 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20743 }
20744 }
20745
20746 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_gt_8_subtile) {
20747 TEST_REQUIRES_ARM_NEON;
20748 for (size_t k = 9; k < 16; k++) {
20749 for (uint32_t n = 1; n <= 16; n++) {
20750 for (uint32_t m = 1; m <= 4; m++) {
20751 GemmMicrokernelTester()
20752 .mr(4)
20753 .nr(16)
20754 .kr(1)
20755 .sr(1)
20756 .m(m)
20757 .n(n)
20758 .k(k)
20759 .iterations(1)
20760 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20761 }
20762 }
20763 }
20764 }
20765
20766 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_div_8) {
20767 TEST_REQUIRES_ARM_NEON;
20768 for (size_t k = 16; k <= 80; k += 8) {
20769 GemmMicrokernelTester()
20770 .mr(4)
20771 .nr(16)
20772 .kr(1)
20773 .sr(1)
20774 .m(4)
20775 .n(16)
20776 .k(k)
20777 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20778 }
20779 }
20780
20781 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, k_div_8_subtile) {
20782 TEST_REQUIRES_ARM_NEON;
20783 for (size_t k = 16; k <= 80; k += 8) {
20784 for (uint32_t n = 1; n <= 16; n++) {
20785 for (uint32_t m = 1; m <= 4; m++) {
20786 GemmMicrokernelTester()
20787 .mr(4)
20788 .nr(16)
20789 .kr(1)
20790 .sr(1)
20791 .m(m)
20792 .n(n)
20793 .k(k)
20794 .iterations(1)
20795 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20796 }
20797 }
20798 }
20799 }
20800
20801 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_gt_16) {
20802 TEST_REQUIRES_ARM_NEON;
20803 for (uint32_t n = 17; n < 32; n++) {
20804 for (size_t k = 1; k <= 40; k += 9) {
20805 GemmMicrokernelTester()
20806 .mr(4)
20807 .nr(16)
20808 .kr(1)
20809 .sr(1)
20810 .m(4)
20811 .n(n)
20812 .k(k)
20813 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20814 }
20815 }
20816 }
20817
20818 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_gt_16_strided_cn) {
20819 TEST_REQUIRES_ARM_NEON;
20820 for (uint32_t n = 17; n < 32; n++) {
20821 for (size_t k = 1; k <= 40; k += 9) {
20822 GemmMicrokernelTester()
20823 .mr(4)
20824 .nr(16)
20825 .kr(1)
20826 .sr(1)
20827 .m(4)
20828 .n(n)
20829 .k(k)
20830 .cn_stride(19)
20831 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20832 }
20833 }
20834 }
20835
20836 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_gt_16_subtile) {
20837 TEST_REQUIRES_ARM_NEON;
20838 for (uint32_t n = 17; n < 32; n++) {
20839 for (size_t k = 1; k <= 40; k += 9) {
20840 for (uint32_t m = 1; m <= 4; m++) {
20841 GemmMicrokernelTester()
20842 .mr(4)
20843 .nr(16)
20844 .kr(1)
20845 .sr(1)
20846 .m(m)
20847 .n(n)
20848 .k(k)
20849 .iterations(1)
20850 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20851 }
20852 }
20853 }
20854 }
20855
20856 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_div_16) {
20857 TEST_REQUIRES_ARM_NEON;
20858 for (uint32_t n = 32; n <= 48; n += 16) {
20859 for (size_t k = 1; k <= 40; k += 9) {
20860 GemmMicrokernelTester()
20861 .mr(4)
20862 .nr(16)
20863 .kr(1)
20864 .sr(1)
20865 .m(4)
20866 .n(n)
20867 .k(k)
20868 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20869 }
20870 }
20871 }
20872
20873 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_div_16_strided_cn) {
20874 TEST_REQUIRES_ARM_NEON;
20875 for (uint32_t n = 32; n <= 48; n += 16) {
20876 for (size_t k = 1; k <= 40; k += 9) {
20877 GemmMicrokernelTester()
20878 .mr(4)
20879 .nr(16)
20880 .kr(1)
20881 .sr(1)
20882 .m(4)
20883 .n(n)
20884 .k(k)
20885 .cn_stride(19)
20886 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20887 }
20888 }
20889 }
20890
20891 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_div_16_subtile) {
20892 TEST_REQUIRES_ARM_NEON;
20893 for (uint32_t n = 32; n <= 48; n += 16) {
20894 for (size_t k = 1; k <= 40; k += 9) {
20895 for (uint32_t m = 1; m <= 4; m++) {
20896 GemmMicrokernelTester()
20897 .mr(4)
20898 .nr(16)
20899 .kr(1)
20900 .sr(1)
20901 .m(m)
20902 .n(n)
20903 .k(k)
20904 .iterations(1)
20905 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20906 }
20907 }
20908 }
20909 }
20910
20911 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, small_kernel) {
20912 TEST_REQUIRES_ARM_NEON;
20913 for (size_t k = 1; k <= 40; k += 9) {
20914 GemmMicrokernelTester()
20915 .mr(4)
20916 .nr(16)
20917 .kr(1)
20918 .sr(1)
20919 .m(4)
20920 .n(16)
20921 .k(k)
20922 .ks(3)
20923 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20924 }
20925 }
20926
20927 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, small_kernel_subtile) {
20928 TEST_REQUIRES_ARM_NEON;
20929 for (size_t k = 1; k <= 40; k += 9) {
20930 for (uint32_t n = 1; n <= 16; n++) {
20931 for (uint32_t m = 1; m <= 4; m++) {
20932 GemmMicrokernelTester()
20933 .mr(4)
20934 .nr(16)
20935 .kr(1)
20936 .sr(1)
20937 .m(m)
20938 .n(n)
20939 .k(k)
20940 .ks(3)
20941 .iterations(1)
20942 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20943 }
20944 }
20945 }
20946 }
20947
20948 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_gt_16_small_kernel) {
20949 TEST_REQUIRES_ARM_NEON;
20950 for (uint32_t n = 17; n < 32; n++) {
20951 for (size_t k = 1; k <= 40; k += 9) {
20952 GemmMicrokernelTester()
20953 .mr(4)
20954 .nr(16)
20955 .kr(1)
20956 .sr(1)
20957 .m(4)
20958 .n(n)
20959 .k(k)
20960 .ks(3)
20961 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20962 }
20963 }
20964 }
20965
20966 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, n_div_16_small_kernel) {
20967 TEST_REQUIRES_ARM_NEON;
20968 for (uint32_t n = 32; n <= 48; n += 16) {
20969 for (size_t k = 1; k <= 40; k += 9) {
20970 GemmMicrokernelTester()
20971 .mr(4)
20972 .nr(16)
20973 .kr(1)
20974 .sr(1)
20975 .m(4)
20976 .n(n)
20977 .k(k)
20978 .ks(3)
20979 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
20980 }
20981 }
20982 }
20983
20984 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, strided_cm_subtile) {
20985 TEST_REQUIRES_ARM_NEON;
20986 for (size_t k = 1; k <= 40; k += 9) {
20987 for (uint32_t n = 1; n <= 16; n++) {
20988 for (uint32_t m = 1; m <= 4; m++) {
20989 GemmMicrokernelTester()
20990 .mr(4)
20991 .nr(16)
20992 .kr(1)
20993 .sr(1)
20994 .m(m)
20995 .n(n)
20996 .k(k)
20997 .cm_stride(19)
20998 .iterations(1)
20999 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21000 }
21001 }
21002 }
21003 }
21004
21005 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, a_offset) {
21006 TEST_REQUIRES_ARM_NEON;
21007 for (size_t k = 1; k <= 40; k += 9) {
21008 GemmMicrokernelTester()
21009 .mr(4)
21010 .nr(16)
21011 .kr(1)
21012 .sr(1)
21013 .m(4)
21014 .n(16)
21015 .k(k)
21016 .ks(3)
21017 .a_offset(163)
21018 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21019 }
21020 }
21021
21022 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, zero) {
21023 TEST_REQUIRES_ARM_NEON;
21024 for (size_t k = 1; k <= 40; k += 9) {
21025 for (uint32_t mz = 0; mz < 4; mz++) {
21026 GemmMicrokernelTester()
21027 .mr(4)
21028 .nr(16)
21029 .kr(1)
21030 .sr(1)
21031 .m(4)
21032 .n(16)
21033 .k(k)
21034 .ks(3)
21035 .a_offset(163)
21036 .zero_index(mz)
21037 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21038 }
21039 }
21040 }
21041
21042 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, qmin) {
21043 TEST_REQUIRES_ARM_NEON;
21044 GemmMicrokernelTester()
21045 .mr(4)
21046 .nr(16)
21047 .kr(1)
21048 .sr(1)
21049 .m(4)
21050 .n(16)
21051 .k(8)
21052 .qmin(128)
21053 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21054 }
21055
21056 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, qmax) {
21057 TEST_REQUIRES_ARM_NEON;
21058 GemmMicrokernelTester()
21059 .mr(4)
21060 .nr(16)
21061 .kr(1)
21062 .sr(1)
21063 .m(4)
21064 .n(16)
21065 .k(8)
21066 .qmax(128)
21067 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21068 }
21069
21070 TEST(QS8_IGEMM_MINMAX_RNDNU_4X16__NEON_MLAL_LANE, strided_cm) {
21071 TEST_REQUIRES_ARM_NEON;
21072 GemmMicrokernelTester()
21073 .mr(4)
21074 .nr(16)
21075 .kr(1)
21076 .sr(1)
21077 .m(4)
21078 .n(16)
21079 .k(8)
21080 .cm_stride(19)
21081 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_4x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21082 }
21083#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
21084
21085
21086#if XNN_ARCH_ARM || XNN_ARCH_ARM64
21087 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8) {
21088 TEST_REQUIRES_ARM_NEON;
21089 GemmMicrokernelTester()
21090 .mr(6)
21091 .nr(16)
21092 .kr(1)
21093 .sr(1)
21094 .m(6)
21095 .n(16)
21096 .k(8)
21097 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21098 }
21099
21100 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, strided_cn) {
21101 TEST_REQUIRES_ARM_NEON;
21102 GemmMicrokernelTester()
21103 .mr(6)
21104 .nr(16)
21105 .kr(1)
21106 .sr(1)
21107 .m(6)
21108 .n(16)
21109 .k(8)
21110 .cn_stride(19)
21111 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21112 }
21113
21114 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8_subtile) {
21115 TEST_REQUIRES_ARM_NEON;
21116 for (uint32_t n = 1; n <= 16; n++) {
21117 for (uint32_t m = 1; m <= 6; m++) {
21118 GemmMicrokernelTester()
21119 .mr(6)
21120 .nr(16)
21121 .kr(1)
21122 .sr(1)
21123 .m(m)
21124 .n(n)
21125 .k(8)
21126 .iterations(1)
21127 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21128 }
21129 }
21130 }
21131
21132 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8_subtile_m) {
21133 TEST_REQUIRES_ARM_NEON;
21134 for (uint32_t m = 1; m <= 6; m++) {
21135 GemmMicrokernelTester()
21136 .mr(6)
21137 .nr(16)
21138 .kr(1)
21139 .sr(1)
21140 .m(m)
21141 .n(16)
21142 .k(8)
21143 .iterations(1)
21144 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21145 }
21146 }
21147
21148 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_eq_8_subtile_n) {
21149 TEST_REQUIRES_ARM_NEON;
21150 for (uint32_t n = 1; n <= 16; n++) {
21151 GemmMicrokernelTester()
21152 .mr(6)
21153 .nr(16)
21154 .kr(1)
21155 .sr(1)
21156 .m(6)
21157 .n(n)
21158 .k(8)
21159 .iterations(1)
21160 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21161 }
21162 }
21163
21164 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_lt_8) {
21165 TEST_REQUIRES_ARM_NEON;
21166 for (size_t k = 1; k < 8; k++) {
21167 GemmMicrokernelTester()
21168 .mr(6)
21169 .nr(16)
21170 .kr(1)
21171 .sr(1)
21172 .m(6)
21173 .n(16)
21174 .k(k)
21175 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21176 }
21177 }
21178
21179 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_lt_8_subtile) {
21180 TEST_REQUIRES_ARM_NEON;
21181 for (size_t k = 1; k < 8; k++) {
21182 for (uint32_t n = 1; n <= 16; n++) {
21183 for (uint32_t m = 1; m <= 6; m++) {
21184 GemmMicrokernelTester()
21185 .mr(6)
21186 .nr(16)
21187 .kr(1)
21188 .sr(1)
21189 .m(m)
21190 .n(n)
21191 .k(k)
21192 .iterations(1)
21193 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21194 }
21195 }
21196 }
21197 }
21198
21199 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_gt_8) {
21200 TEST_REQUIRES_ARM_NEON;
21201 for (size_t k = 9; k < 16; k++) {
21202 GemmMicrokernelTester()
21203 .mr(6)
21204 .nr(16)
21205 .kr(1)
21206 .sr(1)
21207 .m(6)
21208 .n(16)
21209 .k(k)
21210 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21211 }
21212 }
21213
21214 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_gt_8_subtile) {
21215 TEST_REQUIRES_ARM_NEON;
21216 for (size_t k = 9; k < 16; k++) {
21217 for (uint32_t n = 1; n <= 16; n++) {
21218 for (uint32_t m = 1; m <= 6; m++) {
21219 GemmMicrokernelTester()
21220 .mr(6)
21221 .nr(16)
21222 .kr(1)
21223 .sr(1)
21224 .m(m)
21225 .n(n)
21226 .k(k)
21227 .iterations(1)
21228 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21229 }
21230 }
21231 }
21232 }
21233
21234 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_div_8) {
21235 TEST_REQUIRES_ARM_NEON;
21236 for (size_t k = 16; k <= 80; k += 8) {
21237 GemmMicrokernelTester()
21238 .mr(6)
21239 .nr(16)
21240 .kr(1)
21241 .sr(1)
21242 .m(6)
21243 .n(16)
21244 .k(k)
21245 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21246 }
21247 }
21248
21249 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, k_div_8_subtile) {
21250 TEST_REQUIRES_ARM_NEON;
21251 for (size_t k = 16; k <= 80; k += 8) {
21252 for (uint32_t n = 1; n <= 16; n++) {
21253 for (uint32_t m = 1; m <= 6; m++) {
21254 GemmMicrokernelTester()
21255 .mr(6)
21256 .nr(16)
21257 .kr(1)
21258 .sr(1)
21259 .m(m)
21260 .n(n)
21261 .k(k)
21262 .iterations(1)
21263 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21264 }
21265 }
21266 }
21267 }
21268
21269 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16) {
21270 TEST_REQUIRES_ARM_NEON;
21271 for (uint32_t n = 17; n < 32; n++) {
21272 for (size_t k = 1; k <= 40; k += 9) {
21273 GemmMicrokernelTester()
21274 .mr(6)
21275 .nr(16)
21276 .kr(1)
21277 .sr(1)
21278 .m(6)
21279 .n(n)
21280 .k(k)
21281 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21282 }
21283 }
21284 }
21285
21286 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16_strided_cn) {
21287 TEST_REQUIRES_ARM_NEON;
21288 for (uint32_t n = 17; n < 32; n++) {
21289 for (size_t k = 1; k <= 40; k += 9) {
21290 GemmMicrokernelTester()
21291 .mr(6)
21292 .nr(16)
21293 .kr(1)
21294 .sr(1)
21295 .m(6)
21296 .n(n)
21297 .k(k)
21298 .cn_stride(19)
21299 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21300 }
21301 }
21302 }
21303
21304 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16_subtile) {
21305 TEST_REQUIRES_ARM_NEON;
21306 for (uint32_t n = 17; n < 32; n++) {
21307 for (size_t k = 1; k <= 40; k += 9) {
21308 for (uint32_t m = 1; m <= 6; m++) {
21309 GemmMicrokernelTester()
21310 .mr(6)
21311 .nr(16)
21312 .kr(1)
21313 .sr(1)
21314 .m(m)
21315 .n(n)
21316 .k(k)
21317 .iterations(1)
21318 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21319 }
21320 }
21321 }
21322 }
21323
21324 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16) {
21325 TEST_REQUIRES_ARM_NEON;
21326 for (uint32_t n = 32; n <= 48; n += 16) {
21327 for (size_t k = 1; k <= 40; k += 9) {
21328 GemmMicrokernelTester()
21329 .mr(6)
21330 .nr(16)
21331 .kr(1)
21332 .sr(1)
21333 .m(6)
21334 .n(n)
21335 .k(k)
21336 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21337 }
21338 }
21339 }
21340
21341 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16_strided_cn) {
21342 TEST_REQUIRES_ARM_NEON;
21343 for (uint32_t n = 32; n <= 48; n += 16) {
21344 for (size_t k = 1; k <= 40; k += 9) {
21345 GemmMicrokernelTester()
21346 .mr(6)
21347 .nr(16)
21348 .kr(1)
21349 .sr(1)
21350 .m(6)
21351 .n(n)
21352 .k(k)
21353 .cn_stride(19)
21354 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21355 }
21356 }
21357 }
21358
21359 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16_subtile) {
21360 TEST_REQUIRES_ARM_NEON;
21361 for (uint32_t n = 32; n <= 48; n += 16) {
21362 for (size_t k = 1; k <= 40; k += 9) {
21363 for (uint32_t m = 1; m <= 6; m++) {
21364 GemmMicrokernelTester()
21365 .mr(6)
21366 .nr(16)
21367 .kr(1)
21368 .sr(1)
21369 .m(m)
21370 .n(n)
21371 .k(k)
21372 .iterations(1)
21373 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21374 }
21375 }
21376 }
21377 }
21378
21379 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, small_kernel) {
21380 TEST_REQUIRES_ARM_NEON;
21381 for (size_t k = 1; k <= 40; k += 9) {
21382 GemmMicrokernelTester()
21383 .mr(6)
21384 .nr(16)
21385 .kr(1)
21386 .sr(1)
21387 .m(6)
21388 .n(16)
21389 .k(k)
21390 .ks(3)
21391 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21392 }
21393 }
21394
21395 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, small_kernel_subtile) {
21396 TEST_REQUIRES_ARM_NEON;
21397 for (size_t k = 1; k <= 40; k += 9) {
21398 for (uint32_t n = 1; n <= 16; n++) {
21399 for (uint32_t m = 1; m <= 6; m++) {
21400 GemmMicrokernelTester()
21401 .mr(6)
21402 .nr(16)
21403 .kr(1)
21404 .sr(1)
21405 .m(m)
21406 .n(n)
21407 .k(k)
21408 .ks(3)
21409 .iterations(1)
21410 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21411 }
21412 }
21413 }
21414 }
21415
21416 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_gt_16_small_kernel) {
21417 TEST_REQUIRES_ARM_NEON;
21418 for (uint32_t n = 17; n < 32; n++) {
21419 for (size_t k = 1; k <= 40; k += 9) {
21420 GemmMicrokernelTester()
21421 .mr(6)
21422 .nr(16)
21423 .kr(1)
21424 .sr(1)
21425 .m(6)
21426 .n(n)
21427 .k(k)
21428 .ks(3)
21429 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21430 }
21431 }
21432 }
21433
21434 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, n_div_16_small_kernel) {
21435 TEST_REQUIRES_ARM_NEON;
21436 for (uint32_t n = 32; n <= 48; n += 16) {
21437 for (size_t k = 1; k <= 40; k += 9) {
21438 GemmMicrokernelTester()
21439 .mr(6)
21440 .nr(16)
21441 .kr(1)
21442 .sr(1)
21443 .m(6)
21444 .n(n)
21445 .k(k)
21446 .ks(3)
21447 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21448 }
21449 }
21450 }
21451
21452 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, strided_cm_subtile) {
21453 TEST_REQUIRES_ARM_NEON;
21454 for (size_t k = 1; k <= 40; k += 9) {
21455 for (uint32_t n = 1; n <= 16; n++) {
21456 for (uint32_t m = 1; m <= 6; m++) {
21457 GemmMicrokernelTester()
21458 .mr(6)
21459 .nr(16)
21460 .kr(1)
21461 .sr(1)
21462 .m(m)
21463 .n(n)
21464 .k(k)
21465 .cm_stride(19)
21466 .iterations(1)
21467 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21468 }
21469 }
21470 }
21471 }
21472
21473 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, a_offset) {
21474 TEST_REQUIRES_ARM_NEON;
21475 for (size_t k = 1; k <= 40; k += 9) {
21476 GemmMicrokernelTester()
21477 .mr(6)
21478 .nr(16)
21479 .kr(1)
21480 .sr(1)
21481 .m(6)
21482 .n(16)
21483 .k(k)
21484 .ks(3)
21485 .a_offset(251)
21486 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21487 }
21488 }
21489
21490 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, zero) {
21491 TEST_REQUIRES_ARM_NEON;
21492 for (size_t k = 1; k <= 40; k += 9) {
21493 for (uint32_t mz = 0; mz < 6; mz++) {
21494 GemmMicrokernelTester()
21495 .mr(6)
21496 .nr(16)
21497 .kr(1)
21498 .sr(1)
21499 .m(6)
21500 .n(16)
21501 .k(k)
21502 .ks(3)
21503 .a_offset(251)
21504 .zero_index(mz)
21505 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21506 }
21507 }
21508 }
21509
21510 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, qmin) {
21511 TEST_REQUIRES_ARM_NEON;
21512 GemmMicrokernelTester()
21513 .mr(6)
21514 .nr(16)
21515 .kr(1)
21516 .sr(1)
21517 .m(6)
21518 .n(16)
21519 .k(8)
21520 .qmin(128)
21521 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21522 }
21523
21524 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, qmax) {
21525 TEST_REQUIRES_ARM_NEON;
21526 GemmMicrokernelTester()
21527 .mr(6)
21528 .nr(16)
21529 .kr(1)
21530 .sr(1)
21531 .m(6)
21532 .n(16)
21533 .k(8)
21534 .qmax(128)
21535 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21536 }
21537
21538 TEST(QS8_IGEMM_MINMAX_RNDNU_6X16__NEON_MLAL_LANE, strided_cm) {
21539 TEST_REQUIRES_ARM_NEON;
21540 GemmMicrokernelTester()
21541 .mr(6)
21542 .nr(16)
21543 .kr(1)
21544 .sr(1)
21545 .m(6)
21546 .n(16)
21547 .k(8)
21548 .cm_stride(19)
21549 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_6x16__neon_mlal_lane, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21550 }
21551#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
21552
21553
21554#if XNN_ARCH_ARM || XNN_ARCH_ARM64
21555 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8) {
21556 TEST_REQUIRES_ARM_NEON;
21557 GemmMicrokernelTester()
21558 .mr(2)
21559 .nr(16)
21560 .kr(1)
21561 .sr(1)
21562 .m(2)
21563 .n(16)
21564 .k(8)
21565 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21566 }
21567
21568 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, strided_cn) {
21569 TEST_REQUIRES_ARM_NEON;
21570 GemmMicrokernelTester()
21571 .mr(2)
21572 .nr(16)
21573 .kr(1)
21574 .sr(1)
21575 .m(2)
21576 .n(16)
21577 .k(8)
21578 .cn_stride(19)
21579 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21580 }
21581
21582 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile) {
21583 TEST_REQUIRES_ARM_NEON;
21584 for (uint32_t n = 1; n <= 16; n++) {
21585 for (uint32_t m = 1; m <= 2; m++) {
21586 GemmMicrokernelTester()
21587 .mr(2)
21588 .nr(16)
21589 .kr(1)
21590 .sr(1)
21591 .m(m)
21592 .n(n)
21593 .k(8)
21594 .iterations(1)
21595 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21596 }
21597 }
21598 }
21599
21600 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_m) {
21601 TEST_REQUIRES_ARM_NEON;
21602 for (uint32_t m = 1; m <= 2; m++) {
21603 GemmMicrokernelTester()
21604 .mr(2)
21605 .nr(16)
21606 .kr(1)
21607 .sr(1)
21608 .m(m)
21609 .n(16)
21610 .k(8)
21611 .iterations(1)
21612 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21613 }
21614 }
21615
21616 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_eq_8_subtile_n) {
21617 TEST_REQUIRES_ARM_NEON;
21618 for (uint32_t n = 1; n <= 16; n++) {
21619 GemmMicrokernelTester()
21620 .mr(2)
21621 .nr(16)
21622 .kr(1)
21623 .sr(1)
21624 .m(2)
21625 .n(n)
21626 .k(8)
21627 .iterations(1)
21628 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21629 }
21630 }
21631
21632 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_lt_8) {
21633 TEST_REQUIRES_ARM_NEON;
21634 for (size_t k = 1; k < 8; k++) {
21635 GemmMicrokernelTester()
21636 .mr(2)
21637 .nr(16)
21638 .kr(1)
21639 .sr(1)
21640 .m(2)
21641 .n(16)
21642 .k(k)
21643 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21644 }
21645 }
21646
21647 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_lt_8_subtile) {
21648 TEST_REQUIRES_ARM_NEON;
21649 for (size_t k = 1; k < 8; k++) {
21650 for (uint32_t n = 1; n <= 16; n++) {
21651 for (uint32_t m = 1; m <= 2; m++) {
21652 GemmMicrokernelTester()
21653 .mr(2)
21654 .nr(16)
21655 .kr(1)
21656 .sr(1)
21657 .m(m)
21658 .n(n)
21659 .k(k)
21660 .iterations(1)
21661 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21662 }
21663 }
21664 }
21665 }
21666
21667 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_gt_8) {
21668 TEST_REQUIRES_ARM_NEON;
21669 for (size_t k = 9; k < 16; k++) {
21670 GemmMicrokernelTester()
21671 .mr(2)
21672 .nr(16)
21673 .kr(1)
21674 .sr(1)
21675 .m(2)
21676 .n(16)
21677 .k(k)
21678 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21679 }
21680 }
21681
21682 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_gt_8_subtile) {
21683 TEST_REQUIRES_ARM_NEON;
21684 for (size_t k = 9; k < 16; k++) {
21685 for (uint32_t n = 1; n <= 16; n++) {
21686 for (uint32_t m = 1; m <= 2; m++) {
21687 GemmMicrokernelTester()
21688 .mr(2)
21689 .nr(16)
21690 .kr(1)
21691 .sr(1)
21692 .m(m)
21693 .n(n)
21694 .k(k)
21695 .iterations(1)
21696 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21697 }
21698 }
21699 }
21700 }
21701
21702 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_div_8) {
21703 TEST_REQUIRES_ARM_NEON;
21704 for (size_t k = 16; k <= 80; k += 8) {
21705 GemmMicrokernelTester()
21706 .mr(2)
21707 .nr(16)
21708 .kr(1)
21709 .sr(1)
21710 .m(2)
21711 .n(16)
21712 .k(k)
21713 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21714 }
21715 }
21716
21717 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, k_div_8_subtile) {
21718 TEST_REQUIRES_ARM_NEON;
21719 for (size_t k = 16; k <= 80; k += 8) {
21720 for (uint32_t n = 1; n <= 16; n++) {
21721 for (uint32_t m = 1; m <= 2; m++) {
21722 GemmMicrokernelTester()
21723 .mr(2)
21724 .nr(16)
21725 .kr(1)
21726 .sr(1)
21727 .m(m)
21728 .n(n)
21729 .k(k)
21730 .iterations(1)
21731 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21732 }
21733 }
21734 }
21735 }
21736
21737 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16) {
21738 TEST_REQUIRES_ARM_NEON;
21739 for (uint32_t n = 17; n < 32; n++) {
21740 for (size_t k = 1; k <= 40; k += 9) {
21741 GemmMicrokernelTester()
21742 .mr(2)
21743 .nr(16)
21744 .kr(1)
21745 .sr(1)
21746 .m(2)
21747 .n(n)
21748 .k(k)
21749 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21750 }
21751 }
21752 }
21753
21754 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16_strided_cn) {
21755 TEST_REQUIRES_ARM_NEON;
21756 for (uint32_t n = 17; n < 32; n++) {
21757 for (size_t k = 1; k <= 40; k += 9) {
21758 GemmMicrokernelTester()
21759 .mr(2)
21760 .nr(16)
21761 .kr(1)
21762 .sr(1)
21763 .m(2)
21764 .n(n)
21765 .k(k)
21766 .cn_stride(19)
21767 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21768 }
21769 }
21770 }
21771
21772 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16_subtile) {
21773 TEST_REQUIRES_ARM_NEON;
21774 for (uint32_t n = 17; n < 32; n++) {
21775 for (size_t k = 1; k <= 40; k += 9) {
21776 for (uint32_t m = 1; m <= 2; m++) {
21777 GemmMicrokernelTester()
21778 .mr(2)
21779 .nr(16)
21780 .kr(1)
21781 .sr(1)
21782 .m(m)
21783 .n(n)
21784 .k(k)
21785 .iterations(1)
21786 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21787 }
21788 }
21789 }
21790 }
21791
21792 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16) {
21793 TEST_REQUIRES_ARM_NEON;
21794 for (uint32_t n = 32; n <= 48; n += 16) {
21795 for (size_t k = 1; k <= 40; k += 9) {
21796 GemmMicrokernelTester()
21797 .mr(2)
21798 .nr(16)
21799 .kr(1)
21800 .sr(1)
21801 .m(2)
21802 .n(n)
21803 .k(k)
21804 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21805 }
21806 }
21807 }
21808
21809 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16_strided_cn) {
21810 TEST_REQUIRES_ARM_NEON;
21811 for (uint32_t n = 32; n <= 48; n += 16) {
21812 for (size_t k = 1; k <= 40; k += 9) {
21813 GemmMicrokernelTester()
21814 .mr(2)
21815 .nr(16)
21816 .kr(1)
21817 .sr(1)
21818 .m(2)
21819 .n(n)
21820 .k(k)
21821 .cn_stride(19)
21822 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21823 }
21824 }
21825 }
21826
21827 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16_subtile) {
21828 TEST_REQUIRES_ARM_NEON;
21829 for (uint32_t n = 32; n <= 48; n += 16) {
21830 for (size_t k = 1; k <= 40; k += 9) {
21831 for (uint32_t m = 1; m <= 2; m++) {
21832 GemmMicrokernelTester()
21833 .mr(2)
21834 .nr(16)
21835 .kr(1)
21836 .sr(1)
21837 .m(m)
21838 .n(n)
21839 .k(k)
21840 .iterations(1)
21841 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21842 }
21843 }
21844 }
21845 }
21846
21847 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, small_kernel) {
21848 TEST_REQUIRES_ARM_NEON;
21849 for (size_t k = 1; k <= 40; k += 9) {
21850 GemmMicrokernelTester()
21851 .mr(2)
21852 .nr(16)
21853 .kr(1)
21854 .sr(1)
21855 .m(2)
21856 .n(16)
21857 .k(k)
21858 .ks(3)
21859 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21860 }
21861 }
21862
21863 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, small_kernel_subtile) {
21864 TEST_REQUIRES_ARM_NEON;
21865 for (size_t k = 1; k <= 40; k += 9) {
21866 for (uint32_t n = 1; n <= 16; n++) {
21867 for (uint32_t m = 1; m <= 2; m++) {
21868 GemmMicrokernelTester()
21869 .mr(2)
21870 .nr(16)
21871 .kr(1)
21872 .sr(1)
21873 .m(m)
21874 .n(n)
21875 .k(k)
21876 .ks(3)
21877 .iterations(1)
21878 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21879 }
21880 }
21881 }
21882 }
21883
21884 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_gt_16_small_kernel) {
21885 TEST_REQUIRES_ARM_NEON;
21886 for (uint32_t n = 17; n < 32; n++) {
21887 for (size_t k = 1; k <= 40; k += 9) {
21888 GemmMicrokernelTester()
21889 .mr(2)
21890 .nr(16)
21891 .kr(1)
21892 .sr(1)
21893 .m(2)
21894 .n(n)
21895 .k(k)
21896 .ks(3)
21897 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21898 }
21899 }
21900 }
21901
21902 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, n_div_16_small_kernel) {
21903 TEST_REQUIRES_ARM_NEON;
21904 for (uint32_t n = 32; n <= 48; n += 16) {
21905 for (size_t k = 1; k <= 40; k += 9) {
21906 GemmMicrokernelTester()
21907 .mr(2)
21908 .nr(16)
21909 .kr(1)
21910 .sr(1)
21911 .m(2)
21912 .n(n)
21913 .k(k)
21914 .ks(3)
21915 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21916 }
21917 }
21918 }
21919
21920 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, strided_cm_subtile) {
21921 TEST_REQUIRES_ARM_NEON;
21922 for (size_t k = 1; k <= 40; k += 9) {
21923 for (uint32_t n = 1; n <= 16; n++) {
21924 for (uint32_t m = 1; m <= 2; m++) {
21925 GemmMicrokernelTester()
21926 .mr(2)
21927 .nr(16)
21928 .kr(1)
21929 .sr(1)
21930 .m(m)
21931 .n(n)
21932 .k(k)
21933 .cm_stride(19)
21934 .iterations(1)
21935 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21936 }
21937 }
21938 }
21939 }
21940
21941 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, a_offset) {
21942 TEST_REQUIRES_ARM_NEON;
21943 for (size_t k = 1; k <= 40; k += 9) {
21944 GemmMicrokernelTester()
21945 .mr(2)
21946 .nr(16)
21947 .kr(1)
21948 .sr(1)
21949 .m(2)
21950 .n(16)
21951 .k(k)
21952 .ks(3)
21953 .a_offset(83)
21954 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21955 }
21956 }
21957
21958 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, zero) {
21959 TEST_REQUIRES_ARM_NEON;
21960 for (size_t k = 1; k <= 40; k += 9) {
21961 for (uint32_t mz = 0; mz < 2; mz++) {
21962 GemmMicrokernelTester()
21963 .mr(2)
21964 .nr(16)
21965 .kr(1)
21966 .sr(1)
21967 .m(2)
21968 .n(16)
21969 .k(k)
21970 .ks(3)
21971 .a_offset(83)
21972 .zero_index(mz)
21973 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21974 }
21975 }
21976 }
21977
21978 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, qmin) {
21979 TEST_REQUIRES_ARM_NEON;
21980 GemmMicrokernelTester()
21981 .mr(2)
21982 .nr(16)
21983 .kr(1)
21984 .sr(1)
21985 .m(2)
21986 .n(16)
21987 .k(8)
21988 .qmin(128)
21989 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
21990 }
21991
21992 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, qmax) {
21993 TEST_REQUIRES_ARM_NEON;
21994 GemmMicrokernelTester()
21995 .mr(2)
21996 .nr(16)
21997 .kr(1)
21998 .sr(1)
21999 .m(2)
22000 .n(16)
22001 .k(8)
22002 .qmax(128)
22003 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22004 }
22005
22006 TEST(QS8_IGEMM_MINMAX_RNDNU_2X16__NEON_MLAL_LANE_PRFM, strided_cm) {
22007 TEST_REQUIRES_ARM_NEON;
22008 GemmMicrokernelTester()
22009 .mr(2)
22010 .nr(16)
22011 .kr(1)
22012 .sr(1)
22013 .m(2)
22014 .n(16)
22015 .k(8)
22016 .cm_stride(19)
22017 .Test(xnn_qs8_igemm_minmax_rndnu_ukernel_2x16__neon_mlal_lane_prfm, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22018 }
22019#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
22020
22021
22022#if XNN_ARCH_ARM && !XNN_PLATFORM_IOS && XNN_ENABLE_ASSEMBLY && XNN_PLATFORM_JIT
22023 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8) {
22024 TEST_REQUIRES_ARM_NEON_DOT;
22025 GemmMicrokernelTester()
22026 .mr(4)
22027 .nr(8)
22028 .kr(4)
22029 .sr(1)
22030 .m(4)
22031 .n(8)
22032 .k(8)
22033 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22034 }
22035
22036 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, strided_cn) {
22037 TEST_REQUIRES_ARM_NEON_DOT;
22038 GemmMicrokernelTester()
22039 .mr(4)
22040 .nr(8)
22041 .kr(4)
22042 .sr(1)
22043 .m(4)
22044 .n(8)
22045 .k(8)
22046 .cn_stride(11)
22047 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22048 }
22049
22050 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8_subtile) {
22051 TEST_REQUIRES_ARM_NEON_DOT;
22052 for (uint32_t n = 1; n <= 8; n++) {
22053 for (uint32_t m = 1; m <= 4; m++) {
22054 GemmMicrokernelTester()
22055 .mr(4)
22056 .nr(8)
22057 .kr(4)
22058 .sr(1)
22059 .m(m)
22060 .n(n)
22061 .k(8)
22062 .iterations(1)
22063 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22064 }
22065 }
22066 }
22067
22068 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8_subtile_m) {
22069 TEST_REQUIRES_ARM_NEON_DOT;
22070 for (uint32_t m = 1; m <= 4; m++) {
22071 GemmMicrokernelTester()
22072 .mr(4)
22073 .nr(8)
22074 .kr(4)
22075 .sr(1)
22076 .m(m)
22077 .n(8)
22078 .k(8)
22079 .iterations(1)
22080 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22081 }
22082 }
22083
22084 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_eq_8_subtile_n) {
22085 TEST_REQUIRES_ARM_NEON_DOT;
22086 for (uint32_t n = 1; n <= 8; n++) {
22087 GemmMicrokernelTester()
22088 .mr(4)
22089 .nr(8)
22090 .kr(4)
22091 .sr(1)
22092 .m(4)
22093 .n(n)
22094 .k(8)
22095 .iterations(1)
22096 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22097 }
22098 }
22099
22100 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_lt_8) {
22101 TEST_REQUIRES_ARM_NEON_DOT;
22102 for (size_t k = 1; k < 8; k++) {
22103 GemmMicrokernelTester()
22104 .mr(4)
22105 .nr(8)
22106 .kr(4)
22107 .sr(1)
22108 .m(4)
22109 .n(8)
22110 .k(k)
22111 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22112 }
22113 }
22114
22115 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_lt_8_subtile) {
22116 TEST_REQUIRES_ARM_NEON_DOT;
22117 for (size_t k = 1; k < 8; k++) {
22118 for (uint32_t n = 1; n <= 8; n++) {
22119 for (uint32_t m = 1; m <= 4; m++) {
22120 GemmMicrokernelTester()
22121 .mr(4)
22122 .nr(8)
22123 .kr(4)
22124 .sr(1)
22125 .m(m)
22126 .n(n)
22127 .k(k)
22128 .iterations(1)
22129 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22130 }
22131 }
22132 }
22133 }
22134
22135 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_gt_8) {
22136 TEST_REQUIRES_ARM_NEON_DOT;
22137 for (size_t k = 9; k < 16; k++) {
22138 GemmMicrokernelTester()
22139 .mr(4)
22140 .nr(8)
22141 .kr(4)
22142 .sr(1)
22143 .m(4)
22144 .n(8)
22145 .k(k)
22146 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22147 }
22148 }
22149
22150 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_gt_8_subtile) {
22151 TEST_REQUIRES_ARM_NEON_DOT;
22152 for (size_t k = 9; k < 16; k++) {
22153 for (uint32_t n = 1; n <= 8; n++) {
22154 for (uint32_t m = 1; m <= 4; m++) {
22155 GemmMicrokernelTester()
22156 .mr(4)
22157 .nr(8)
22158 .kr(4)
22159 .sr(1)
22160 .m(m)
22161 .n(n)
22162 .k(k)
22163 .iterations(1)
22164 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22165 }
22166 }
22167 }
22168 }
22169
22170 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_div_8) {
22171 TEST_REQUIRES_ARM_NEON_DOT;
22172 for (size_t k = 16; k <= 80; k += 8) {
22173 GemmMicrokernelTester()
22174 .mr(4)
22175 .nr(8)
22176 .kr(4)
22177 .sr(1)
22178 .m(4)
22179 .n(8)
22180 .k(k)
22181 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22182 }
22183 }
22184
22185 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, k_div_8_subtile) {
22186 TEST_REQUIRES_ARM_NEON_DOT;
22187 for (size_t k = 16; k <= 80; k += 8) {
22188 for (uint32_t n = 1; n <= 8; n++) {
22189 for (uint32_t m = 1; m <= 4; m++) {
22190 GemmMicrokernelTester()
22191 .mr(4)
22192 .nr(8)
22193 .kr(4)
22194 .sr(1)
22195 .m(m)
22196 .n(n)
22197 .k(k)
22198 .iterations(1)
22199 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22200 }
22201 }
22202 }
22203 }
22204
22205 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8) {
22206 TEST_REQUIRES_ARM_NEON_DOT;
22207 for (uint32_t n = 9; n < 16; n++) {
22208 for (size_t k = 1; k <= 40; k += 9) {
22209 GemmMicrokernelTester()
22210 .mr(4)
22211 .nr(8)
22212 .kr(4)
22213 .sr(1)
22214 .m(4)
22215 .n(n)
22216 .k(k)
22217 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22218 }
22219 }
22220 }
22221
22222 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8_strided_cn) {
22223 TEST_REQUIRES_ARM_NEON_DOT;
22224 for (uint32_t n = 9; n < 16; n++) {
22225 for (size_t k = 1; k <= 40; k += 9) {
22226 GemmMicrokernelTester()
22227 .mr(4)
22228 .nr(8)
22229 .kr(4)
22230 .sr(1)
22231 .m(4)
22232 .n(n)
22233 .k(k)
22234 .cn_stride(11)
22235 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22236 }
22237 }
22238 }
22239
22240 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8_subtile) {
22241 TEST_REQUIRES_ARM_NEON_DOT;
22242 for (uint32_t n = 9; n < 16; n++) {
22243 for (size_t k = 1; k <= 40; k += 9) {
22244 for (uint32_t m = 1; m <= 4; m++) {
22245 GemmMicrokernelTester()
22246 .mr(4)
22247 .nr(8)
22248 .kr(4)
22249 .sr(1)
22250 .m(m)
22251 .n(n)
22252 .k(k)
22253 .iterations(1)
22254 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22255 }
22256 }
22257 }
22258 }
22259
22260 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8) {
22261 TEST_REQUIRES_ARM_NEON_DOT;
22262 for (uint32_t n = 16; n <= 24; n += 8) {
22263 for (size_t k = 1; k <= 40; k += 9) {
22264 GemmMicrokernelTester()
22265 .mr(4)
22266 .nr(8)
22267 .kr(4)
22268 .sr(1)
22269 .m(4)
22270 .n(n)
22271 .k(k)
22272 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22273 }
22274 }
22275 }
22276
22277 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8_strided_cn) {
22278 TEST_REQUIRES_ARM_NEON_DOT;
22279 for (uint32_t n = 16; n <= 24; n += 8) {
22280 for (size_t k = 1; k <= 40; k += 9) {
22281 GemmMicrokernelTester()
22282 .mr(4)
22283 .nr(8)
22284 .kr(4)
22285 .sr(1)
22286 .m(4)
22287 .n(n)
22288 .k(k)
22289 .cn_stride(11)
22290 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22291 }
22292 }
22293 }
22294
22295 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8_subtile) {
22296 TEST_REQUIRES_ARM_NEON_DOT;
22297 for (uint32_t n = 16; n <= 24; n += 8) {
22298 for (size_t k = 1; k <= 40; k += 9) {
22299 for (uint32_t m = 1; m <= 4; m++) {
22300 GemmMicrokernelTester()
22301 .mr(4)
22302 .nr(8)
22303 .kr(4)
22304 .sr(1)
22305 .m(m)
22306 .n(n)
22307 .k(k)
22308 .iterations(1)
22309 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22310 }
22311 }
22312 }
22313 }
22314
22315 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, small_kernel) {
22316 TEST_REQUIRES_ARM_NEON_DOT;
22317 for (size_t k = 1; k <= 40; k += 9) {
22318 GemmMicrokernelTester()
22319 .mr(4)
22320 .nr(8)
22321 .kr(4)
22322 .sr(1)
22323 .m(4)
22324 .n(8)
22325 .k(k)
22326 .ks(3)
22327 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22328 }
22329 }
22330
22331 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, small_kernel_subtile) {
22332 TEST_REQUIRES_ARM_NEON_DOT;
22333 for (size_t k = 1; k <= 40; k += 9) {
22334 for (uint32_t n = 1; n <= 8; n++) {
22335 for (uint32_t m = 1; m <= 4; m++) {
22336 GemmMicrokernelTester()
22337 .mr(4)
22338 .nr(8)
22339 .kr(4)
22340 .sr(1)
22341 .m(m)
22342 .n(n)
22343 .k(k)
22344 .ks(3)
22345 .iterations(1)
22346 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22347 }
22348 }
22349 }
22350 }
22351
22352 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_gt_8_small_kernel) {
22353 TEST_REQUIRES_ARM_NEON_DOT;
22354 for (uint32_t n = 9; n < 16; n++) {
22355 for (size_t k = 1; k <= 40; k += 9) {
22356 GemmMicrokernelTester()
22357 .mr(4)
22358 .nr(8)
22359 .kr(4)
22360 .sr(1)
22361 .m(4)
22362 .n(n)
22363 .k(k)
22364 .ks(3)
22365 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22366 }
22367 }
22368 }
22369
22370 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, n_div_8_small_kernel) {
22371 TEST_REQUIRES_ARM_NEON_DOT;
22372 for (uint32_t n = 16; n <= 24; n += 8) {
22373 for (size_t k = 1; k <= 40; k += 9) {
22374 GemmMicrokernelTester()
22375 .mr(4)
22376 .nr(8)
22377 .kr(4)
22378 .sr(1)
22379 .m(4)
22380 .n(n)
22381 .k(k)
22382 .ks(3)
22383 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22384 }
22385 }
22386 }
22387
22388 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, strided_cm_subtile) {
22389 TEST_REQUIRES_ARM_NEON_DOT;
22390 for (size_t k = 1; k <= 40; k += 9) {
22391 for (uint32_t n = 1; n <= 8; n++) {
22392 for (uint32_t m = 1; m <= 4; m++) {
22393 GemmMicrokernelTester()
22394 .mr(4)
22395 .nr(8)
22396 .kr(4)
22397 .sr(1)
22398 .m(m)
22399 .n(n)
22400 .k(k)
22401 .cm_stride(11)
22402 .iterations(1)
22403 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22404 }
22405 }
22406 }
22407 }
22408
22409 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, a_offset) {
22410 TEST_REQUIRES_ARM_NEON_DOT;
22411 for (size_t k = 1; k <= 40; k += 9) {
22412 GemmMicrokernelTester()
22413 .mr(4)
22414 .nr(8)
22415 .kr(4)
22416 .sr(1)
22417 .m(4)
22418 .n(8)
22419 .k(k)
22420 .ks(3)
22421 .a_offset(163)
22422 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22423 }
22424 }
22425
22426 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, zero) {
22427 TEST_REQUIRES_ARM_NEON_DOT;
22428 for (size_t k = 1; k <= 40; k += 9) {
22429 for (uint32_t mz = 0; mz < 4; mz++) {
22430 GemmMicrokernelTester()
22431 .mr(4)
22432 .nr(8)
22433 .kr(4)
22434 .sr(1)
22435 .m(4)
22436 .n(8)
22437 .k(k)
22438 .ks(3)
22439 .a_offset(163)
22440 .zero_index(mz)
22441 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22442 }
22443 }
22444 }
22445
22446 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, qmin) {
22447 TEST_REQUIRES_ARM_NEON_DOT;
22448 GemmMicrokernelTester()
22449 .mr(4)
22450 .nr(8)
22451 .kr(4)
22452 .sr(1)
22453 .m(4)
22454 .n(8)
22455 .k(8)
22456 .qmin(128)
22457 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22458 }
22459
22460 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, qmax) {
22461 TEST_REQUIRES_ARM_NEON_DOT;
22462 GemmMicrokernelTester()
22463 .mr(4)
22464 .nr(8)
22465 .kr(4)
22466 .sr(1)
22467 .m(4)
22468 .n(8)
22469 .k(8)
22470 .qmax(128)
22471 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22472 }
22473
22474 TEST(GENERATE_QS8_IGEMM_RNDNU_4X8C4__AARCH32_NEONDOT_LD64, strided_cm) {
22475 TEST_REQUIRES_ARM_NEON_DOT;
22476 GemmMicrokernelTester()
22477 .mr(4)
22478 .nr(8)
22479 .kr(4)
22480 .sr(1)
22481 .m(4)
22482 .n(8)
22483 .k(8)
22484 .cm_stride(11)
22485 .Test(xnn_generate_qs8_igemm_rndnu_ukernel_4x8c4__aarch32_neondot_ld64, xnn_init_qs8_conv_minmax_rndnu_neon_params, xnn_qs8_requantize_rndnu);
22486 }
22487#endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS && XNN_ENABLE_ASSEMBLY && XNN_PLATFORM_JIT